{ "updated": "2026-11-02", "about": "存算一体芯片图谱：每家公司 / 研究线的技术路线（where 存内 / 近存 / 相邻 / 上游，mode 模拟 / 数字 / 混合，media 存储介质）、工艺节点（process_node，只写公开的）、量产状态（stage：量产 / 样片出货 / 流片 / 研究 / 已停，带日期和出处）、目标场景（scene：端侧 / 边缘 / 数据中心），以及融资、估值和关键数据，每个数字附出处（src）。layer 对应 layers.json 的 code。product_line_of = 大公司的产品线 / 研究线。", "companies": [ { "name": "Irréversible", "layer": "a1", "url": "https://irreversible.tech/", "hq": "加拿大舍布鲁克", "region": "海外", "founded": 2024, "desc": "忆阻器全模拟存内推理芯片，面向机器人和工业边缘", "where": "存内", "mode": "模拟", "media": [ "RRAM" ], "stage": { "v": "研究", "date": "2026-01", "note": "预种子轮阶段，资金用于芯片迭代", "src": "https://cyclemomentum.com/irreversible-making-artificial-intelligence-compatible-with-the-planets-physical-limits-2/" }, "scene": [ "边缘" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2026-01", "round": "预种子轮", "amount": "未披露", "lead": "Quantacet（UCeed、Frostbite Capital、Cycle Momentum、Tofino Capital、BDC、Capital Labs 参与）", "src": "https://cyclemomentum.com/irreversible-making-artificial-intelligence-compatible-with-the-planets-physical-limits-2/" } ], "note": "总部城市官方稿未写；成立年份来自 PitchBook（线索级）；介质按忆阻器记 RRAM，官方稿未明确器件类型，待核。官网未查到。" }, { "name": "Mentium Technologies", "layer": "a1", "url": "https://www.mentium.tech", "hq": "戈莱塔", "region": "海外", "desc": "忆阻器模拟存内计算加数字加速器的混合 AI 协处理器", "where": "存内", "mode": "混合", "media": [ "RRAM" ], "stage": { "v": "研究", "note": "NASA SBIR 二期已做出并测试硅片，三期目标做出抗辐射 TRL 7 原型", "src": "https://techport.nasa.gov/projects/154327" }, "scene": [ "边缘" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2026-06", "round": "私募（SEC Form D）", "amount": "已售约 $7.5M（拟募约 $8.8M）", "lead": "未披露", "src": "https://aum13f.com/fund/mentium-technologies-inc" } ], "customers": [ "NASA（SBIR 二期，合同额 $4.57M）" ], "note": "融资来自 SEC Form D 的镜像页，权威度一般；成立年份官方未写；与应用材料（Applied Materials）合作忆阻器工艺（NASA 页面）。" }, { "name": "TetraMem", "layer": "a1", "url": "https://tetramem.com", "hq": "圣何塞", "region": "海外", "desc": "多级 RRAM 阵列做模拟存内计算 SoC", "where": "存内", "mode": "模拟", "media": [ "RRAM" ], "process_node": { "v": "22nm", "src": "https://secure.businesswire.com/news/home/20260516556464/en/TetraMem-Announces-22nm-Multi-Level-RRAM-Analog-In-Memory-Computing-SoC-Milestone" }, "stage": { "v": "流片", "date": "2026-05", "note": "2026-05 公告完成流片、生产及首批硅验证", "src": "https://secure.businesswire.com/news/home/20260516556464/en/TetraMem-Announces-22nm-Multi-Level-RRAM-Analog-In-Memory-Computing-SoC-Milestone" }, "scene": [ "端侧", "边缘" ], "total_raised": "累计融资未披露", "note": "仅用官网与官方新闻稿；成立年份、融资额官方渠道未公布，故不填。" }, { "name": "新忆科技 Innostar-Mem", "layer": "a1", "url": "https://www.innomemtech.com/home", "hq": "北京", "region": "中国", "founded": 2018, "founded_month": "2018-06", "desc": "清华系 RRAM 厂商，布局 RRAM 存算 NPU/LPU 推理芯片", "where": "存内", "mode": "模拟", "media": [ "RRAM" ], "process_node": { "v": "40nm 至 22nm", "src": "https://scc.pku.edu.cn/frontpage/pku/html/recruitmentinfoForm.html?positionDetailId=60b0b7bebcfa4b3f802046c9dd8ec3fd&" }, "stage": { "v": "量产", "date": "2021", "note": "仅指独立式 RRAM 存储器（累计出货超 3000 万颗）；存算推理芯片仍在开发，已用验证芯片测能效", "src": "https://www.innomemtech.com/home" }, "scene": [ "端侧", "边缘", "数据中心" ], "total_raised": "累计融资过亿元（公司招聘页口径；投资方含中关村科学城、SK 海力士）", "note": "各轮金额与日期未找到一手出处（B 轮 2024 仅见公众号）；mode 暂记模拟，官方未明示，待核；存算芯片流片、节点未查到。验证芯片说法见 https://m.10jqka.com.cn/20261012/c680590072.shtml 。" }, { "name": "昕原半导体 Innostar", "layer": "a1", "url": "https://www.innostar-semi.com/", "hq": "上海", "region": "中国", "founded": 2019, "founded_month": "2019-10", "desc": "自建 ReRAM 中试线，提供 ReRAM 存算一体 IP 与芯片", "where": "存内", "mode": "模拟", "media": [ "RRAM" ], "process_node": { "v": "28/22nm（ReRAM 中试线口径）", "src": "https://m.x-techcon.com/article/88372.html" }, "stage": { "v": "样片出货", "date": "2025-12", "note": "公司称存储芯片、存算 IP 已商用出货交付；存算一体 IP 另有高校 IP 授权加 MPW 流片合同", "src": "https://m.x-techcon.com/article/88372.html" }, "scene": [ "端侧", "边缘", "数据中心" ], "total_raised": "累计披露融资约 9.78 亿元（投资界口径，10 次）", "funding": [ { "date": "2026-07", "round": "C+++ 轮", "amount": "未披露", "lead": "未披露（参与：云锋基金、上海自贸区基金、联新资本、联升资本等）", "src": "https://stock.stockstar.com/RB2026072100039001.shtml" }, { "date": "2025-12", "round": "C+ 轮", "amount": "未披露", "lead": "云锋基金、瑞壹投资", "src": "https://vc.pedaily.cn/company/185094.html" } ], "customers": [ "东北师范大学（ReRAM 存算 IP 授权加流片，成交 199.8 万元，2024-11）" ], "note": "ReRAM 存算路线是模拟还是数字，官方口径未写明，mode 按 ReRAM 存内通行做法暂记模拟，待核；估值未查到；股东含字节、蚂蚁、红杉中国等，来源为工商变更报道（x-techcon 2025-12-25；成交公告 http://zgggzy.com/jieguo/show.php?itemid=32495 ）。" }, { "name": "梦忆科技 MemInTech", "layer": "a1", "url": "https://memintech.net/", "hq": "武汉", "region": "中国", "founded": 2026, "founded_month": "2026-03", "desc": "华中科大忆阻器团队孵化，做多介质超异构存算芯片", "where": "存内", "mode": "模拟", "media": [ "RRAM" ], "stage": { "v": "研究", "date": "2026-09", "note": "种子轮资金用于核心技术攻关与原型芯片开发验证", "src": "https://stcn.com/article/detail/4188071.html" }, "scene": [ "端侧" ], "total_raised": "累计融资千万级（种子轮）", "funding": [ { "date": "2026-09", "round": "种子轮", "amount": "千万级", "lead": "华工投资（独家，华工科技全资子公司）", "src": "https://stcn.com/article/detail/4188071.html" } ], "note": "2026-03-20 在武汉光谷成立；华中科大拟以 120 万元转让 4 项存算专利（2026-09 公示，见 https://www.toutiao.com/article/7686766382217855532/ ）；mode 暂记模拟，官方称忆阻器等多介质，待核；官网未查到。" }, { "name": "铭芯启睿 MemSilicon", "layer": "a1", "url": "https://memsilicon.com/", "hq": "北京", "region": "中国", "founded": 2024, "founded_month": "2024-05", "desc": "源自中科院团队的 RRAM 存算，在研大模型推理 LPU", "where": "存内", "mode": "混合", "media": [ "RRAM" ], "stage": { "v": "流片", "date": "2026-01", "note": "已完成产品工程批验证流片", "src": "https://m.tmtpost.com/7830791.html" }, "scene": [ "端侧", "数据中心" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2026-10", "round": "A 轮", "amount": "超 4 亿元", "lead": "博原资本（博世创投、创领资本、米哈游、中关村科学城公司、兆易创新跟投）", "src": "https://cbgc.scol.com.cn/news/8072068" }, { "date": "2026-01", "round": "Pre-A 轮", "amount": "超 1 亿元", "lead": "国开科创、联想创投（中芯聚源、顺禧基金、恒裕投资跟投）", "src": "https://m.tmtpost.com/7830791.html" } ], "note": "mode 记混合是因 2026-10 稿称做 RRAM 加 LPU 的存算协同，模拟/数字比例未写明，待核；url 暂用报道页（官网未查到）；估值未查到。" }, { "name": "闪易半导体 Flashbillion", "layer": "a1", "url": "http://www.flashbillion.com/", "hq": "上海", "region": "中国", "founded": 2017, "founded_month": "2017-07", "desc": "PLRAM 忆阻器阵列存算一体芯片", "where": "存内", "mode": "模拟", "media": [ "其他" ], "stage": { "v": "流片", "date": "2019-10", "note": "发布首款存算一体芯片（6MB 模拟存储，8bit/单元）；之后进展未查到", "src": "https://baike.baidu.com/item/%E6%9D%AD%E5%B7%9E%E9%97%AA%E4%BA%BF%E5%8D%8A%E5%AF%BC%E4%BD%93%E6%9C%89%E9%99%90%E5%85%AC%E5%8F%B8" }, "scene": [ "端侧" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2021-04", "round": "A 轮", "amount": "未披露", "lead": "上海张江科技创业投资（襄创创业孵化器参与）", "src": "https://pitchhub.36kr.com/project/1817519684011908" } ], "note": "资料少。曾用名杭州闪亿半导体；介质为 PLRAM 型忆阻器（非闪存，名字易混）。2026-01 京东旗下公司入股、90nm 产线可量产等说法仅见公众号，未找到一手出处，不采。节点、量产状态、客户未查到。" }, { "name": "Anaflash", "layer": "a2", "url": "https://www.anaflash.com", "hq": "森尼韦尔", "region": "海外", "founded": 2017, "desc": "标准逻辑兼容嵌入式闪存加存内计算的 AI MCU", "where": "存内", "mode": "混合", "media": [ "其他" ], "process_node": { "v": "28nm", "src": "https://www.storagenewsletter.com/2025/11/10/anaflash-advances-embedded-flash-memory-for-next-gen-smart-edge-devices-with-samsung-foundry/" }, "stage": { "v": "样片出货", "date": "2025-11", "note": "28nm AI MCU 原型向合作伙伴和客户送样", "src": "https://www.theguru.co.kr/news/article.html?no=93832" }, "scene": [ "端侧", "边缘" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2024-11", "round": "未披露（A 轮仍在进行）", "amount": "未披露", "lead": "Stonebridge Ventures（L&S Venture Capital、Lotte Ventures、Plug and Play 跟投）", "src": "https://investor.wedbush.com/wedbush/article/bizwire-2024-11-19-anaflash-secures-new-funding-to-advance-edge-ai-processing-and-innovation-in-partnership-with-leading-global-ventures" } ], "note": "嵌入式闪存为 4 bit/cell 的 Logic-eFlash，是否严格 NOR 型官方稿未写，故介质记其他；2025-02 收购 Legato Logic 的时域存内计算技术（官网新闻 https://www.anaflash.com/news ）；韩国子公司 SEMIBRAIN。" }, { "name": "Mythic", "layer": "a2", "url": "https://www.mythic.ai", "hq": "帕洛阿尔托", "region": "海外", "founded": 2012, "desc": "闪存单元模拟存内计算 APU，重建架构后覆盖边缘到数据中心", "where": "存内", "mode": "模拟", "media": [ "NOR Flash" ], "scene": [ "边缘", "数据中心" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2025-12", "round": "未披露轮次名", "amount": "$125M", "lead": "DCVC（NEA、Atreides、Future Ventures、SBVA、S3 Ventures、Honda、Lockheed Martin 等跟投）", "src": "https://www.s3vc.com/news-and-insights/mythic-to-challenge-ais-gpu-pantheon-with-100x-energy-advantage-and-oversubscribed-125m-raise" } ], "key_data": [ { "k": "能效（公司称，下一代 APU）", "v": "120 TOPS/W，采用 SST memBrain 闪存单元 IP", "date": "2026-03", "src": "https://www.epdtonthenet.net/article/221058/" } ], "customers": [ "Honda（合作开发车规 AI 芯片）" ], "note": "2026-05 收购 videantis（德国数字处理器 IP 公司，价格未披露，见 https://ecapital.vc/news/mythic-acquires-videantis-%e2%80%92-ecapital-remains-shareholder/ ）。新一代 APU 的流片与出货状态、工艺节点未查到；融资轮次名各来源不一，故不写。" }, { "name": "优存科技 UCUN", "layer": "a2", "url": "http://www.ucuntech.com/", "hq": "南京", "region": "中国", "founded": 2018, "founded_month": "2018-05", "desc": "NOR Flash 芯片厂商，同时做基于 NOR 架构的存内计算芯片", "where": "存内", "mode": "模拟", "media": [ "NOR Flash" ], "scene": [ "端侧" ], "valuation": { "v": "投前估值约 5.90 亿元", "date": "2026-08", "kind": "国民技术增资公告披露的投前估值", "src": "https://www.cnfin.com/announ/detail/index.html?id=841610732106&code=300077&dannoun=lcdetail&announ=lc" }, "total_raised": "累计融资未披露", "funding": [ { "date": "2026-09", "round": "新一轮（公告未写轮次）", "amount": "国民技术出资 6000 万元（占 8.3333%），另有其他投资方", "lead": "国民技术（上市公司）", "post": "投前估值 5.90 亿元", "src": "https://www.cnfin.com/announ/detail/index.html?id=841610732106&code=300077&dannoun=lcdetail&announ=lc" } ], "note": "主业是 NOR Flash 芯片，存内计算为在研产品线，进度、节点、模拟/数字路线均未查到（mode 暂记模拟，待核）；2025-06 成都科创投集团入股见 36 氪项目页。" }, { "name": "知存科技 Witmem", "layer": "a2", "url": "https://www.witmem.com", "hq": "北京", "region": "中国", "founded": 2017, "founded_month": "2017-10", "desc": "NOR Flash 模拟存内计算芯片，已量产用于耳机等可穿戴", "where": "存内", "mode": "模拟", "media": [ "NOR Flash" ], "stage": { "v": "量产", "date": "2022", "note": "WTM2101 于 2022 年批量生产；2026-05 与安克联合开发的 Thus 存算音频芯片发布", "src": "http://www.sme.buaa.edu.cn/__local/0/FD/5B/57E0E0B8F524DA90CDF5B81E10A_950FCEB4_37512.pdf" }, "scene": [ "端侧" ], "total_raised": "累计融资近 8 亿元（截至 2023-01，百度百科口径）", "funding": [ { "date": "2023-01", "round": "B2 轮", "amount": "2 亿元", "lead": "国投创业（水木春锦资本、领航新界跟投）", "src": "https://www.pinhui.wang/191408.shtml" } ], "key_data": [ { "k": "WTM2101 算力与功耗", "v": ">50 GOPS；5 µA–3 mA", "src": "https://en.witmem.com/WTM2101.php" } ], "customers": [ "安克创新（联合开发 Thus 存算音频芯片）", "维迈通 VIMOTO（骑行头盔耳机）" ], "note": "工艺节点 40nm 仅见知乎综述，未找到一手出处，故不填；2023-01 之后是否有新融资未查到；估值未查到。安克合作见 http://www.c114.net.cn/industry/87921.html 。" }, { "name": "AIStorm", "layer": "a3", "url": "https://aistorm.ai", "hq": "休斯敦", "region": "海外", "founded": 2018, "desc": "电荷域模拟神经元，做传感器内 AI 成像芯片", "where": "相邻", "mode": "模拟", "media": [ "电容" ], "stage": { "v": "样片出货", "date": "2025-08", "note": "Cheetah HS（260K FPS 带首层 AI 的成像芯片）2025-08 发布并可经销商购买", "src": "https://www.sec.gov/Archives/edgar/data/928876/000117891325002760/zk2533618.htm" }, "scene": [ "端侧", "边缘" ], "total_raised": "累计融资约 $29.2M（Jon Peddie 口径）", "funding": [ { "date": "2020-12", "round": "B 轮", "amount": "$16M", "lead": "未披露（AsusTek、Egis、Knowles、Meyer、Senvest 等参与）", "src": "https://www.jonpeddie.com/news/aistorm-puts-ai-inside-the-image-sensor/" }, { "date": "2019-02", "round": "A 轮", "amount": "$13.2M", "lead": "未披露", "src": "https://www.jonpeddie.com/news/aistorm-puts-ai-inside-the-image-sensor/" } ], "key_data": [ { "k": "电荷域边缘模拟 AI 能效（公司称）", "v": "最高 600 TOPS/W（40nm）", "src": "https://aistorm.ai/" } ], "customers": [ "Tower Semiconductor（联合发布 Cheetah HS）" ], "note": "where 记相邻：是传感器内电荷域 AI，不是存储阵列内计算；放 a3 的理由是电荷域模拟，待定。" }, { "name": "Ambient Scientific", "layer": "a3", "url": "https://www.ambientscientific.ai", "hq": "圣克拉拉", "region": "海外", "founded": 2017, "desc": "DigAn 数模混合架构：SRAM 内模拟 MAC 的超低功耗 AI 处理器", "where": "存内", "mode": "混合", "media": [ "SRAM" ], "process_node": { "v": "40nm（首代 GPX-10）", "src": "https://techinsights.com/vendor/ambient-scientific" }, "stage": { "v": "样片出货", "date": "2025-10", "note": "GPX10 Pro 送样，计划 2026 Q1 量产；2026-03 已有客户可穿戴产品采用 GPX-10", "src": "https://www.eetindia.co.in/ambient-scientific-building-edge-ai-processors-with-proprietary-tech/" }, "scene": [ "端侧" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2019", "round": "A 轮", "amount": "$17M", "lead": "未披露", "src": "https://techinsights.com/vendor/ambient-scientific" } ], "key_data": [ { "k": "GPX10 Pro 峰值算力与能效（公司称）", "v": "512 GOPS；>7 TOPS/W；常开功耗 <80 µW", "date": "2026", "src": "https://www.ambientscientific.ai/eds" } ], "customers": [ "Dimension NXG（女性安全可穿戴）" ], "note": "2026-03 向 SEC 提交 Form D（拟募 $20M，尚未售出，来源 https://www.tinytechfund.com/company/ambient-scientific-inc ，二手页）；SRAM 内模拟 MAC，介质记 SRAM。" }, { "name": "Areanna AI", "layer": "a3", "url": "https://www.eetimes.com/startup-runs-ai-in-novel-sram/", "hq": "伯克利", "region": "海外", "founded": 2018, "desc": "SRAM 阵列内混合信号矩阵乘，附 ADC/DAC 的存内计算设计", "where": "存内", "mode": "混合", "media": [ "SRAM" ], "stage": { "v": "研究", "date": "2021-04", "note": "2021 年公开架构，仿真阶段；流片与近况未查到", "src": "https://m.ithome.com/html/545215.htm" }, "scene": [ "端侧", "边缘" ], "total_raised": "累计融资约 $22.5 万（仅 NSF SBIR 资助）", "funding": [ { "date": "2019", "round": "NSF SBIR 一期资助", "amount": "$224,996", "lead": "美国国家科学基金会", "src": "https://www.sbir.gov/portfolio/1490539" } ], "note": "资料很少，可能已不活跃；EE Times 2019 报道称仿真能效超 100 TOPS/W（仿真值，非实测）。" }, { "name": "Aspinity", "layer": "a3", "url": "https://www.aspinity.com", "hq": "匹兹堡", "region": "海外", "founded": 2015, "desc": "全模拟机器学习核（AnalogML），让 MCU 在常开感知时休眠", "where": "相邻", "mode": "模拟", "media": [ "其他" ], "stage": { "v": "样片出货", "date": "2020-12", "note": "AML100 评估套件 2020-12 发布；后续量产状态未查到", "src": "https://www.globenewswire.com/en/news-release/2020/12/17/2146970/0/en/Aspinity-Releases-Analog-Voice-First-Evaluation-Kit-for-Battery-Operated-Devices.html" }, "scene": [ "端侧" ], "total_raised": "累计融资超 $19M（公司口径，截至 2023-09）", "funding": [ { "date": "2023-09", "round": "B 轮", "amount": "$5M", "lead": "现有投资方 Anzu Partners、Birchmere Ventures、Mountain State Capital、Riverfront Ventures；新增战略方 Unitrontech", "src": "https://www.businesswire.com/news/home/20230919505548/en/Aspinity-Raises-%245-Million-Series-B-Funding-to-Bring-Breakthrough-AIML-Technology-to-a-Global-Market" } ], "note": "建议 where 记相邻：是模拟域 ML 处理，不是在存储阵列里算；放 a3 或 x2 待定。2023 之后的新融资、节点未查到。" }, { "name": "Blumind", "layer": "a3", "url": "https://blumind.ai", "hq": "渥太华", "region": "海外", "desc": "全模拟晶体管内计算（AMPL）的常开音频 AI 芯片", "where": "存内", "mode": "模拟", "media": [ "其他" ], "stage": { "v": "流片", "note": "BM110 处于原型向量产准备过渡，试点部署计划 2027 年初", "src": "https://fabricinnovation.ca/blumind-inc/" }, "scene": [ "端侧" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2025-04", "round": "A 轮", "amount": "C$20M", "lead": "Cycle Capital、BDC Capital 共同领投（Fusion Fund、Two Small Fish、Real Ventures 跟投）", "src": "https://blumind.ai/blog/blumind-secures-series-a-funding-to-accelerate-analog-ai-revolution/" } ], "key_data": [ { "k": "BM110 常开唤醒功耗", "v": "典型约 30 µA（含麦克风约 50 µA）", "date": "2025-10", "src": "https://blumind.ai/products/bm110/" } ], "note": "自称在晶体管内计算（无 ADC、无 SRAM），是否严格算存内待定；成立年份、节点未查到。2025-09 向首批客户送样的说法仅见聚合站，未采。" }, { "name": "EnCharge AI", "layer": "a3", "url": "https://www.enchargeai.com", "hq": "圣克拉拉", "region": "海外", "founded": 2022, "desc": "电容电荷域模拟存内计算 AI 加速器，先做笔记本客户端", "where": "存内", "mode": "模拟", "media": [ "电容" ], "scene": [ "端侧", "边缘" ], "total_raised": "累计融资超 $144M（公司口径）", "funding": [ { "date": "2025-02", "round": "B 轮", "amount": "$100M+", "lead": "Tiger Global（Maverick Silicon、Samsung Ventures、HH-CTBC 等参与）", "src": "https://www.tradeshownews.com/news/home/20250213156276/en/EnCharge-AI-Closes-Oversubscribed-%24100M-Series-B-Funding-Round-Led-by-Tiger-Global-to-Produce-Transformative-AI-Accelerator-Solutions-with-Unprecedented-Efficiency-for-Client-Computing" } ], "key_data": [ { "k": "EN100 算力与能效（公司称）", "v": "200 TOPS（INT8）；40 TOPS/W 以上", "date": "2025-05", "src": "https://www.eetimes.com/encharge-picks-the-pc-for-its-first-analog-ai-chip/" } ], "note": "EN100 于 2025-05 发布（BusinessWire）；实际出货时间、工艺节点未找到一手出处（B 轮稿称 2025 年商用，法文转载写 2026 年，口径不一）；估值未查到。" }, { "name": "Polyn Semiconductor", "layer": "a3", "url": "https://polyn.ai", "hq": "凯撒利亚", "region": "海外", "founded": 2019, "desc": "模拟神经形态信号处理（NASP）芯片，做语音与振动前端", "where": "相邻", "mode": "模拟", "media": [ "其他" ], "stage": { "v": "流片", "date": "2025-11", "note": "NeuroVoice VAD 为首个硅验证的 NASP 产品；Embedded World 2026 展示", "src": "https://www.allaboutcircuits.com/news/polyn-implements-its-nasp-tech-into-voice-activity-detection-chip/" }, "scene": [ "端侧" ], "total_raised": "累计融资未披露（数据库口径约 $9M，未采）", "key_data": [ { "k": "NeuroVoice VAD 功耗", "v": "30 µW；检测延迟 25 ms", "src": "http://staging8.polyn.ai/wp-content/uploads/2023/10/NeuroVoice-VAD-Product-Brief-v08-5.pdf" } ], "note": "建议 where 记相邻：神经元用运放、突触用薄膜电阻，是模拟网络硬连线，不是存储阵列内计算；融资、节点未查到一手出处，总部以数据库为准。" }, { "name": "Sagence AI", "layer": "a3", "url": "https://www.sagence-ai.com", "hq": "圣克拉拉", "region": "海外", "founded": 2018, "desc": "多值存储单元内做深亚阈值模拟计算的推理 chiplet", "where": "存内", "mode": "模拟", "media": [ "其他" ], "scene": [ "数据中心", "边缘" ], "total_raised": "累计融资约 $58M（出隐身时报道）", "funding": [ { "date": "2024-11", "round": "出隐身（披露累计 $58M）", "amount": "未披露单轮金额", "lead": "未披露（参与：Prosperity7、Khosla Ventures、TDK Ventures、Cambium Capital）", "src": "https://dig.watch/updates/new-startup-tackles-ai-energy-demands-with-analog-tech" } ], "key_data": [ { "k": "对比 GPU（公司称，Llama2-70B）", "v": "功耗低 10 倍、价格低 20 倍、机架空间小 20 倍", "date": "2024-11", "src": "https://www.sagence-ai.com/sagence-ai-emerges-from-stealth-tackling-economic-viability-of-inference-hardware-for-generative-ai/" } ], "note": "原名 Analog Inference。具体存储单元类型（是否闪存）官方稿未写，故介质记其他，若按 a2 闪存放需二轮核实；2024-11 称计划 2025 推出芯片，实际流片/出货状态、节点未查到；Prosperity7 参与见其投资人的 LinkedIn 帖（2024-11）。" }, { "name": "Semron", "layer": "a3", "url": "https://www.semron.ai", "hq": "德累斯顿", "region": "海外", "founded": 2020, "desc": "忆容器 CapRAM 电荷域存内计算，主打 3D 堆叠", "where": "存内", "mode": "模拟", "media": [ "电容" ], "stage": { "v": "研究", "date": "2026-07", "note": "CTO 在 FMS 2026 介绍 3D CapRAM（面向大模型解码），给出技术状态与 3D 集成路线图；未见量产或出货", "src": "https://www.terrapinn.com/conference/future-memory-storage/speaker-kaiuwe-DEMASIUS.stm" }, "scene": [ "端侧", "数据中心" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2024-01", "round": "种子轮", "amount": "€7.3M", "lead": "Join Capital（OTB Ventures、SquareOne Venture Capital、Onsight Ventures 跟投）", "src": "https://tech.eu/2024/01/30/semron-secures-eur73m-seed-funding/" } ], "note": "成立年份各库有 2019 与 2020 两种说法，取 2020；2025 年有欧洲创新理事会（EIC）资助的说法仅见数据库，未找到一手出处。" }, { "name": "Unconventional AI", "layer": "a3", "url": "https://unconv.ai", "hq": "未披露", "region": "海外", "founded": 2025, "desc": "模拟与混合信号概率计算芯片，路线未公开", "where": "相邻", "mode": "模拟", "media": [ "其他" ], "stage": { "v": "研究", "note": "早期；公司称将在数年内试制多种架构原型；官网目前展示的是模拟振荡器系统的软件仿真", "src": "https://mitsloanindia.com/article/unconventional-ai-raises-475-million-seed-round-at-4-5-billion-valuation/" }, "scene": [ "数据中心" ], "valuation": { "v": "$4.5B", "date": "2025-12", "kind": "种子轮投后估值，报道值", "src": "https://www.datacenterdynamics.com/en/news/neuromorphic-compute-startup-unconventional-ai-raises-475m-in-seed-funding/" }, "total_raised": "累计融资约 $475M（种子轮，报道称目标可到 $10 亿）", "funding": [ { "date": "2025-12", "round": "种子轮", "amount": "$475M", "lead": "Andreessen Horowitz、Lightspeed 共同领投", "post": "投后 $4.5B", "src": "https://www.datacenterdynamics.com/en/news/neuromorphic-compute-startup-unconventional-ai-raises-475m-in-seed-funding/" } ], "note": "不是严格存内计算，a16z 称其做模拟与混合信号芯片、存储概率分布；放 a3 还是 x2 类脑待定。" }, { "name": "Articron", "layer": "d1", "url": "https://articron.com", "hq": "首尔", "region": "海外", "founded": 2023, "desc": "SRAM-CIM 的 NPU IP 与 SRAM 设计自动化平台，授权给端侧芯片", "where": "存内", "mode": "数字", "media": [ "SRAM" ], "stage": { "v": "研究", "date": "2026-05", "note": "IP 阶段，客户在评估其 CIM NPU IP", "src": "https://www.thelec.net/news/articleView.html?idxno=14407" }, "scene": [ "端侧", "边缘" ], "funding": [ { "date": "2026-05", "round": "Pre-A 轮", "amount": "30 亿韩元", "lead": "Innopolis Partners 20 亿韩元、HB Investment 10 亿韩元", "src": "https://www.etnews.com/20260511000368?SNS=00002" } ], "note": "卖 IP 与 EDA 工具，不卖芯片，规模很小，图里作长尾代表。公司所在地以 CB Insights 为准在京畿道城南（本条按首尔登记，未能统一）。" }, { "name": "Axelera AI", "layer": "d1", "url": "https://axelera.ai", "hq": "埃因霍温", "region": "海外", "founded": 2021, "desc": "SRAM 数字存内计算（D-IMC）加 RISC-V 的边缘推理芯片 Metis / Europa", "where": "存内", "mode": "数字", "media": [ "SRAM" ], "stage": { "v": "量产", "date": "2026-09", "note": "Metis 卡已商用；二代 Europa 2026-09 发布并随 Dell、Supermicro 服务器交付。公司页未写「量产」一词，按出货口径归入量产", "src": "https://axelera.ai/news/page/1" }, "scene": [ "边缘", "端侧", "数据中心" ], "total_raised": "累计融资未找到公司口径的一手汇总", "funding": [ { "date": "2026-02", "round": "未标轮次", "amount": "超 $250M", "lead": "Innovation Industries 领投（BlackRock、Samsung Catalyst Fund 跟投）", "src": "https://siliconangle.com/2026/02/24/edge-ai-chip-startup-axelera-ai-raises-250m-funding-round/" } ], "key_data": [ { "k": "Metis 单芯片算力 / 功耗", "v": "214 TOPS，约 10W", "date": "2026-02", "src": "https://www.digitaltoday.co.kr/cn/view/28861/axelera-ai-raises-more-than-250-million-for-edge-ai-chips" }, { "k": "Europa 单芯片算力 / 功耗", "v": "629 TOPS，每颗 45W", "date": "2026-09", "src": "https://www.storagereview.com/sv/news/axelera-europa-ships-629-tops-at-45w-per-aipu-in-validated-dell-xe5-and-supermicro-servers" } ], "customers": [ "Dell", "Supermicro", "Prodrive Technologies", "PwC Middle East" ], "note": "工艺节点、估值、营收没查到一手来源。" }, { "name": "GSI Technology", "layer": "d1", "url": "https://www.gsitechnology.com", "hq": "森尼韦尔", "region": "海外", "founded": 1995, "desc": "SRAM 内关联处理器 APU（Gemini-I / II），存内计算用于边缘与向量搜索", "where": "存内", "mode": "数字", "media": [ "SRAM" ], "stage": { "v": "样片出货", "date": "2026-05", "note": "Gemini-II 进入国防无人机 PoC 与首批客户接触；Plato 芯片在设计", "src": "https://ir.gsitechnology.com/news-releases/news-release-details/gsi-technology-inc-announces-fourth-quarter-and-fiscal-year-2026/" }, "scene": [ "边缘" ], "revenue": { "v": "$25.1M", "date": "FY2026（截至 2026-03）", "kind": "全年净营收，主要来自 SRAM 产品", "src": "https://ir.gsitechnology.com/news-releases/news-release-details/gsi-technology-inc-announces-fourth-quarter-and-fiscal-year-2026/" }, "key_data": [ { "k": "Gemini-II 多模态边缘推理（第三方基准）", "v": "首 token 约 3 秒，系统约 30W", "date": "2026-01", "src": "https://ir.gsitechnology.com/news-releases/news-release-details/gsi-technology-inc-announces-third-quarter-fiscal-2026-results" } ], "public": true, "note": "NASDAQ: GSIT。FY2026 期末现金 $67.2M、无负债；另有约 $2M 美国陆军 SBIR 合同。stage 的「样片出货」是对 PoC 阶段的归类，公司稿未用该词。" }, { "name": "九天睿芯 Reexen", "layer": "d1", "url": "https://reexen.com/", "hq": "深圳", "region": "中国", "founded": 2018, "founded_month": "2018-06", "desc": "SRAM 存算一体芯片，在研 SRAM 加 HBF 的 3D 融合存算", "where": "存内", "mode": "数字", "media": [ "SRAM" ], "stage": { "v": "量产", "date": "2026-10", "note": "公司称存算产品已量产销售；SRAM+HBF 3D 融合芯片在研发和流片阶段", "src": "https://www.36kr.com/p/4035227454477189" }, "scene": [ "端侧", "边缘", "数据中心" ], "total_raised": "累计融资未披露（上面是 B 轮半年累计口径）", "funding": [ { "date": "2026-10", "round": "B 轮", "amount": "半年累计 10 亿元", "lead": "未披露领投（美团战投、小鹏汽车星航资本、长存产业基金、元禾璞华、豪威集团先后入局）", "src": "https://wap.stockstar.com/detail/IG2026102200029572" }, { "date": "2026-05", "round": "B2 轮", "amount": "超亿元", "lead": "元禾璞华、哇牛资本等", "src": "https://m.cyzone.cn/article/811172.html" } ], "customers": [ "vivo", "荣耀", "联想" ], "note": "早期有模拟感存算路线，当前主力为 SRAM 数字存算，mode 按当前产品填。估值、营收、工艺节点没查到。" }, { "name": "后摩智能 Houmo", "layer": "d1", "url": "https://www.houmoai.com", "hq": "北京", "region": "中国", "founded": 2020, "founded_month": "2020-11", "desc": "SRAM 数字存算（2D / 3D CIM）端边大模型芯片 M50", "where": "存内", "mode": "数字", "media": [ "SRAM" ], "stage": { "v": "量产", "date": "2026-02", "note": "M50 2025 年初流片，2026-02 量产（行业媒体口径）；官网称 2025-07 发布，另有口径称 2025Q4 量产，时间不一致", "src": "https://www.readaitime.com/news/2026-07-20/1s0q679m" }, "scene": [ "端侧", "边缘" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2025-05", "round": "战略轮", "amount": "未披露", "lead": "中国移动产业链发展基金", "src": "https://www.houmoai.com/AboutUs.html" }, { "date": "2023-06", "round": "A 轮", "amount": "未披露", "lead": "诚通混改、君海创芯共同领投", "src": "https://www.houmoai.com/AboutUs.html" } ], "key_data": [ { "k": "M50 算力 / 功耗", "v": "160 TOPS（INT8），典型 10W，最大 48GB LPDDR5", "date": "2026-10", "src": "https://www.iotdt.com/news/xingyezixun/3148.html" }, { "k": "M50 商业化", "v": "量产不到一年，150+ 定点，客户 100+", "date": "2026-10", "src": "https://www.iotdt.com/news/xingyezixun/3148.html" } ], "customers": [ "联想", "长城电脑" ], "note": "工艺节点没查到一手来源。估值、营收只见低权威页面传言，未采用。官网另有 2024-07 中国移动战略融资口径（百度百科）与官网 2025-05 并存，未能对齐。" }, { "name": "思敏威 CiMicro.Ai", "layer": "d1", "url": "https://cimicro.ai/", "hq": "上海", "region": "中国", "founded": 2024, "desc": "SRAM 存内计算加 3D 近存的「青衡」HPU，端边大模型推理", "where": "存内", "mode": "数字", "media": [ "SRAM" ], "stage": { "v": "研究", "date": "2026-10", "note": "第一代 3D 堆叠芯片「流片在即」，尚未流片", "src": "https://www.cnstock.com/commonDetail/802694" }, "scene": [ "端侧", "边缘" ], "total_raised": "累计两轮，合计约 2 亿元量级（按两次公开口径）", "funding": [ { "date": "2026-10", "round": "天使轮", "amount": "约亿元", "lead": "中兴通讯战略投资，博华产投、松青资本、建信信托、合创资本等跟投", "src": "https://www.zhidx.com/p/600715.html" }, { "date": "2026-05", "round": "种子轮", "amount": "近亿元", "lead": "东方富海、纪源资本领投", "src": "https://media.qimingpian.cn/2026/05/07/400cb2ab4768f908f7493aef87c8bb4e.html" } ], "customers": [ "无问芯穹（商业合作）" ], "note": "成立时间两处口径：上证报 2024-12，企名片 2025，此处取前者。同一公司两轮分别称种子、天使，按媒体原文。" }, { "name": "恒烁半导体 Zbit", "layer": "d1", "url": "https://www.zbitsemi.com", "hq": "合肥", "region": "中国", "founded": 2015, "founded_month": "2015-02", "desc": "NOR Flash 厂商，AI 业务线做 SRAM 数字存算 SoC", "where": "存内", "mode": "数字", "media": [ "SRAM", "NOR Flash" ], "stage": { "v": "研究", "date": "2025-11", "note": "公司答投资者称 NOR Flash 存算一体芯片未量产；2025 年报称 SRAM 存算项目推进商业化，低功耗语音 SoC 设计 2025 下半年启动", "src": "https://dxpress.gelonghui.com/live/2168734" }, "scene": [ "端侧" ], "revenue": { "v": "4.75 亿元", "date": "2025", "kind": "2025 年营业收入，同比 +27.49%，归母净利润 -9,949 万元", "src": "https://ictrade.cn/news/article-details/2518896" }, "total_raised": "2022-08 科创板上市（688416）", "public": true, "note": "建议放数字组：官网 AI 业务线写明是基于 SRAM 的数字存算，NOR Flash 存算仍在研发；因此 a2 不合适。上市公司，融资轮次不适用。" }, { "name": "磐序微电子", "layer": "d1", "url": "https://www.qingdaonews.com/content/2026-08/21/content_23752145.htm", "hq": "青岛", "region": "中国", "founded": 2026, "founded_month": "2026-08", "desc": "家电与工业端侧物理 AI 存算一体芯片，由苹芯技术支撑", "where": "存内", "mode": "数字", "media": [ "SRAM" ], "scene": [ "端侧" ], "customers": [ "家电与工业客户（具体名单未查到）" ], "note": "苹芯科技与国家高端智能化家用电器创新中心共建的合资公司，2026-08-12 落地运营（华兴资本任财务顾问），RC505 获 ICDIA 2026 强芯榜架构创新奖。介质按苹芯技术推定为 SRAM，公司稿未单独写。融资额、量产状态没查到。来源：https://media.qimingpian.cn/2026/08/13/bec5a4c7dece386a48cb69efbb4275b7.html" }, { "name": "苹芯科技 PIMCHIP", "layer": "d1", "url": "https://pimchip.cn", "hq": "北京", "region": "中国", "founded": 2021, "founded_month": "2021-02", "desc": "SRAM 存算一体 NPU / 感知芯片，端侧推理", "where": "存内", "mode": "数字", "media": [ "SRAM" ], "process_node": { "v": "28nm", "src": "https://pimchip.cn/list/9.html" }, "stage": { "v": "流片", "date": "2024-08", "note": "28nm SRAM 存算单元流片点亮；N300 NPU、S300 感知芯片 2024-08 发布；量产出货未查到", "src": "https://baike.baidu.com/item/%E5%8C%97%E4%BA%AC%E8%8B%B9%E8%8A%AF%E7%A7%91%E6%8A%80%E6%9C%89%E9%99%90%E5%85%AC%E5%8F%B8" }, "scene": [ "端侧" ], "total_raised": "累计融资未披露；公开可查最后一轮为 2022-08", "funding": [ { "date": "2022-08", "round": "A 轮", "amount": "千万美元级", "lead": "春华创投领投（红点中国、红杉中国、真格基金跟投）", "src": "https://pimchip.cn/list/9.html" }, { "date": "2021-08", "round": "Pre-A 轮", "amount": "近千万美元", "lead": "红点中国领投", "src": "https://pimchip.cn/list/9.html" } ], "key_data": [ { "k": "无 ADC 架构 SRAM 存内计算引擎能效", "v": "27.38 TOPS/W，面效 1.041 Mb/mm²（28nm）", "date": "2023", "src": "https://pimchip.cn/list/9.html" } ], "note": "2026-08 与国创中心共建磐序微电子（见该条）。估值、营收没查到。" }, { "name": "d-Matrix", "layer": "d2", "url": "https://www.d-matrix.ai", "hq": "圣克拉拉", "region": "海外", "founded": 2019, "desc": "SRAM 数字存内计算加 chiplet 的 Corsair 数据中心推理卡", "where": "存内", "mode": "数字", "media": [ "SRAM" ], "process_node": { "v": "6nm（台积电）", "src": "https://www.servethehome.com/d-matrix-corsair-in-memory-computing-for-ai-inference-at-hot-chips-2025/" }, "stage": { "v": "量产", "date": "2026-06", "note": "公司宣布 Corsair 进入全面量产，开始向超大规模云、新云和前沿实验室批量出货", "src": "https://digital-release.kxan.com/business/press-releases/cision/20260609SF79374/d-matrix-corsair-ai-inference-platform-enters-full-production-to-meet-customer-demand/" }, "scene": [ "数据中心" ], "valuation": { "v": "$2B", "date": "2025-11", "kind": "C 轮投后估值，公司公布", "src": "https://www.datacenterdynamics.com/en/news/chip-startup-d-matrix-raises-275m-against-2bn-valuation/" }, "total_raised": "累计融资约 $450M（公司口径）", "funding": [ { "date": "2025-11", "round": "C 轮", "amount": "$275M", "lead": "Bullhound Capital、Triatomic Capital、Temasek 共同领投（QIA、EDBI、M12、Nautilus、Industry Ventures、Mirae Asset 跟投）", "post": "投后 $2B", "src": "https://www.datacenterdynamics.com/en/news/chip-startup-d-matrix-raises-275m-against-2bn-valuation/" } ], "key_data": [ { "k": "Corsair 能效（Hot Chips 2025 幻灯片）", "v": "38 TOPS/W；8 位 2400 TOPS、4 位 9600 TOPS；2GB SRAM、150 TB/s", "date": "2025-08", "src": "https://www.d-matrix.ai/wp-content/uploads/2025/10/hc2025.dmatrix.SudeepBhoja.v01.pdf" } ], "customers": [ "Gimlet Labs（联合测试：与 GPU 搭配把响应时间从约 24 秒降到 2 秒内）" ], "note": "未查到公开营收。公司称 Corsair 比 GPU 快 10 倍、成本低 3 倍，均为公司口径。" }, { "name": "Fractile", "layer": "d2", "url": "https://www.fractile.ai", "hq": "伦敦", "region": "海外", "founded": 2022, "desc": "计算与存储同片的存内推理芯片，面向大模型数据中心推理", "where": "存内", "mode": "数字", "media": [ "SRAM" ], "scene": [ "数据中心" ], "total_raised": "累计融资未见公司口径汇总", "funding": [ { "date": "2026-05", "round": "B 轮", "amount": "$220M", "lead": "Accel、Factorial Funds、Founders Fund 共同领投", "src": "https://www.datacenterdynamics.com/en/news/fractile-raises-220m-to-accelerate-development-of-ai-inference-chips/" }, { "date": "2024-07", "round": "种子轮", "amount": "$15M", "lead": "Kindred Capital、NATO Innovation Fund、Oxford Science Enterprises 联合领投", "src": "https://thenextweb.com/news/fractile-220m-inference-chip" } ], "note": "B 轮资金用于推进首批推理芯片量产，媒体称尚未量产交付；阶段、工艺节点没查到。有媒体报 Anthropic 在早期洽谈采购（The Information，未证实成交），不列为客户。SRAM 介质来自媒体对其架构的描述（technews.tw 2026-05-03）。" }, { "name": "Rain AI", "layer": "d2", "url": "https://rain.ai", "hq": "旧金山", "region": "海外", "founded": 2017, "desc": "数字存内计算 / 类脑推理芯片，从未出商用芯片，近乎停运", "where": "存内", "mode": "数字", "media": [ "SRAM" ], "stage": { "v": "已停", "date": "2026-08", "note": "收购谈判失败后近乎停运，多数员工离开", "src": "https://www.chaincatcher.com/en/article/2281069" }, "total_raised": "累计融资未披露（第三方库约 $67M，未采）", "funding": [ { "date": "2025-05", "round": "B 轮（未成）", "amount": "计划 $150M", "lead": "私募融资失败，转而探索出售", "src": "https://nypost.com/2025/05/15/business/sam-altman-backed-chipmaking-startup-rain-ai-explores-sale-after-150m-fundraising-flops-sources/" } ], "owner": "2026-08 OpenAI 买下部分专利资产，价格未披露；整体收购谈判未成，公司近乎停运（The Information 报道）", "note": "技术路线（数字存内、SRAM）沿用初稿口径，本轮未找到官方页面核实。" }, { "name": "亿铸科技", "layer": "d2", "url": "https://yizhu-tech.com/", "hq": "苏州", "region": "中国", "founded": 2020, "founded_month": "2020-06", "desc": "全数字通用存算一体（GP-IMC），ReRAM 转向 3D DRAM 的大算力芯片", "where": "存内", "mode": "数字", "media": [ "RRAM", "DRAM" ], "stage": { "v": "流片", "date": "2026-10", "note": "基于 3D DRAM 的全数字存算一体芯片已流片并通过第三方测试，商用芯片测试基本达预设指标", "src": "https://4g.stockstar.com/detail/IG2026102300031276" }, "scene": [ "数据中心" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2026-10", "round": "新一轮", "amount": "超 10 亿元", "lead": "投资方未披露", "src": "https://aichip001.com/p/603963.html" }, { "date": "2025-10", "round": "新一轮", "amount": "未披露", "lead": "兴湘资本、农银国际、盈科值得、合鼎共资本参与", "src": "https://aichip001.com/p/603963.html" }, { "date": "2024-10", "round": "未标轮次", "amount": "数亿元", "lead": "Prosperity7 Ventures 领投（盛视科技、行至资本跟投）", "src": "https://aichip001.com/p/603963.html" } ], "note": "存储介质经历 ReRAM 到 3D DRAM 迭代，此前点亮过忆阻器 POC 芯片。工艺节点、估值、营收没查到。" }, { "name": "千芯科技 TensorChip", "layer": "d2", "url": "https://www.innohere.com/ir/100950/product/101361.html", "hq": "北京", "region": "中国", "founded": 2019, "founded_month": "2019-10", "desc": "可重构存算一体 RMPU 芯片与 IP，面向云端和边缘推理", "where": "存内", "mode": "数字", "media": [ "SRAM" ], "scene": [ "边缘", "数据中心" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2021-03", "round": "未披露轮次", "amount": "数千万元", "lead": "前海融华汇金、NW 投资", "src": "http://news.eeworld.com.cn/xfdz/ic530646.html" } ], "note": "建议改放数字组（d1/d2）：董事长公开称其为 SRAM 数字存算（知乎原文，权威度低，未找到一手确认），不是模拟路线。2024-10 有一轮未披露金额的融资，仅见聚合站，不采。工艺节点、流片/量产状态未查到；官网域名未确认，url 用资料页。" }, { "name": "微纳核芯", "layer": "d2", "url": "https://nanocorechip.com/", "hq": "杭州", "region": "中国", "founded": 2021, "desc": "3D-CIM：DRAM 3D 近存加 SRAM 存算加 RISC-V，端到云大模型推理", "where": "存内", "mode": "混合", "media": [ "SRAM", "DRAM" ], "stage": { "v": "流片", "date": "2026-10", "note": "第一代三颗芯片正从投片向客户验证过渡，计划 2027 年初送测；公司称近期实现规模商业化", "src": "https://awtmt.com/articles/3784665" }, "scene": [ "端侧", "边缘", "数据中心" ], "total_raised": "半年内累计融资 30 亿元（公司口径）", "funding": [ { "date": "2026-10", "round": "C2 轮", "amount": "超 10 亿元", "lead": "建投投资领投（杭州资本、凯泰资本、晓池资本、烟台财金、盛堃投资、国投泰康、领阳基金跟投）", "src": "https://www.dramx.com/News/Memory/20261029-41363.html" }, { "date": "2026-09", "round": "C1 轮", "amount": "10 亿元", "lead": "优势资本、元昊投资、金浦投资、阳光人寿、元禾璞华等（领投未标）", "src": "https://finance.eastmoney.com/a/202609093869638380.html" } ], "key_data": [ { "k": "SRAM 存算对比冯诺依曼（公司实测流片）", "v": "算力密度 4–6 倍，能效 5–10 倍", "date": "2026-09", "src": "https://hznews.hangzhou.com.cn/jingji/content/2026-09/12/content_9305061.htm" } ], "note": "估值未披露。不同媒体对产品状态说法不一：有稿称两条产品线已规模量产销售，有稿称 2027 年初才送测，这里取后者（更保守）。孵化于浙江省北大信息技术高等研究院。" }, { "name": "NEO Semiconductor", "layer": "n1", "url": "https://neosemic.com/neo-semiconductor-announces-the-development-of-its-3d-x-ai-chip-targeted-to-replace-existing-hbm-chips-and-solve-data-bus-bottlenecks/", "hq": "圣何塞", "region": "海外", "founded_month": "2024-08", "desc": "3D X-AI：在 3D DRAM 堆叠中加神经元电路层，目标替代 HBM", "where": "近存", "mode": "数字", "media": [ "DRAM" ], "stage": { "v": "研究", "date": "2026-04", "note": "3D X-AI 为方案公布阶段；其底层 3D X-DRAM 2026-04 公布概念验证（PoC）结果", "src": "https://semiiphub.com/pulse/news/neo-semiconductor-3d-x-dram-investment" }, "scene": [ "数据中心" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2026-04", "round": "战略投资", "amount": "未披露", "lead": "宏碁创办人施振荣领投的一组投资人", "src": "https://semiiphub.com/pulse/news/neo-semiconductor-3d-x-dram-investment" } ], "key_data": [ { "k": "3D X-AI 设计指标", "v": "单颗 300 层 3D DRAM、128Gb，加 1 层 8,000 神经元电路；公司估算单颗 AI 处理吞吐最高 10 TB/s（12 颗堆叠 120 TB/s）", "date": "2024-08", "src": "https://neosemic.com/neo-semiconductor-announces-the-development-of-its-3d-x-ai-chip-targeted-to-replace-existing-hbm-chips-and-solve-data-bus-bottlenecks/" }, { "k": "3D X-DRAM PoC", "v": "读写延迟小于 10ns，85°C 下数据保持超过 1 秒", "date": "2026-04", "src": "https://gfe-desktop.cnyes.com/news/id/6430567" } ], "note": "成立年份没查到。性能与功耗数字（100 倍、99%）是公司估算，不是流片实测。公司计划授权给存储厂量产（Taiwan News 2026-04-24），没有自己的量产芯片。" }, { "name": "NeuroBlade", "layer": "n1", "url": "https://neuroblade.ai/about/", "hq": "特拉维夫", "region": "海外", "founded": 2018, "desc": "SQL 处理单元 SPU，面向数据分析的存内处理加速卡", "where": "近存", "mode": "数字", "stage": { "v": "已停", "date": "2025-10", "note": "核心工程团队并入 AWS Annapurna Labs，公司不再独立运营", "src": "https://www.grovevc.com/news/neuroblade-joins-amazon-web-services-to-drive-the-future-of-data-processing-innovation/" }, "scene": [ "数据中心" ], "total_raised": "累计融资约 $110M", "funding": [ { "date": "2021-10", "round": "B 轮", "amount": "$83M", "lead": "Corner Ventures（EEWorld 转述 Calcalist）", "src": "https://en.eeworld.com.cn/mp/Icbank/a409048.jspx" } ], "key_data": [ { "k": "查询加速", "v": "SPU 对大规模数据查询比领先 CPU 快 4–10 倍（投资方口径）", "date": "2025-10", "src": "https://www.grovevc.com/news/neuroblade-joins-amazon-web-services-to-drive-the-future-of-data-processing-innovation/" } ], "customers": [ "AWS" ], "owner": "2025-10 核心工程团队并入 AWS Annapurna Labs，金额报道称「数千万美元」（Grove），StorageNewsletter 估计低于其融资额", "note": "分层存疑：官方只说 SPU 是 SQL 处理器、有 ASIC/FPGA PCIe 卡形态；2021 年融资稿（Marubeni 公告）称其把数据处理集成在内存里（PIM），但官网近年口径是分析加速卡。建议总编判断是否收，已收则放近存。存储介质没查到，故不写 media。" }, { "name": "Samsung HBM-PIM / LPDDR-PIM", "layer": "n1", "url": "https://news.samsungsemiconductor.com/global/samsung-develops-industrys-first-high-bandwidth-memory-with-ai-processing-power/", "hq": "水原", "region": "海外", "founded": 2021, "founded_month": "2021-02", "desc": "在 DRAM 每个 bank 内置 AI 计算单元：HBM2 Aquabolt-XL，后推出 LPDDR5X-PIM", "where": "近存", "mode": "数字", "media": [ "HBM", "DRAM" ], "stage": { "v": "研究", "date": "2026-08", "note": "HBM-PIM 为 2021 年 PoC 验证；LPDDR5X-PIM 在 Hot Chips 2026 公布架构与规格，2026-02 报道称下半年向客户送样，未见已出货报道", "src": "https://mbiz.heraldcorp.com/article/10675503" }, "scene": [ "数据中心", "端侧" ], "key_data": [ { "k": "HBM-PIM 系统收益", "v": "相对现有 HBM2 系统，性能提升 2 倍以上、系统能耗降低 70% 以上（公司口径）", "date": "2021-02", "src": "https://news.samsungsemiconductor.com/global/samsung-develops-industrys-first-high-bandwidth-memory-with-ai-processing-power/" }, { "k": "HBM-PIM 峰值算力", "v": "1.2 TFLOPS（FP16，每个 bank 一个 300MHz 可编程计算单元）", "date": "2021-02", "src": "https://www.leiphone.com/category/chips/OBlu61efDW1DwHt4.html" }, { "k": "LPDDR5X-PIM 推理", "v": "Llama 3.1 8B：处理时间 5.4 秒 vs 普通 LPDDR5X 12.3 秒；81.3 vs 27 tokens/s；内部带宽 8 倍", "date": "2026-08", "src": "https://ai.cnmo.com/news/816921.html" } ], "product_line_of": "Samsung", "public": true, "note": "另有 CXL-PNM（LPDDR5X 512GB、1.1TB/s，Hot Chips 2023，论文见 https://ieeexplore.ieee.org/abstract/document/10476443）和 AxDIMM，未单列。Samsung 与 SK hynix 在做 LPDDR6-PIM 规范标准化（Herald 2026-02）。LPDDR5X-PIM 的节点、量产时间没查到。stage 选「研究」是因为未见送样出货的报道，请总编定。" }, { "name": "SK hynix AiM / AiMX", "layer": "n1", "url": "https://news.skhynix.com/en/sk-hynix-develops-pim-next-generation-ai-accelerator/", "hq": "利川", "region": "海外", "founded": 2022, "founded_month": "2022-02", "desc": "GDDR6 内置 MAC 的 AiM 存储芯片及多颗 AiM 组成的 AiMX 加速卡", "where": "近存", "mode": "数字", "media": [ "DRAM" ], "stage": { "v": "样片出货", "date": "2022-02", "note": "GDDR6-AiM 样品 2022-02 公布；AiMX 为加速卡原型（2 颗 FPGA + 16 颗 AiM），CES 2026 继续演示；未见量产公告", "src": "https://news.skhynix.com/en/sk-hynix-develops-pim-next-generation-ai-accelerator/" }, "scene": [ "数据中心", "端侧" ], "key_data": [ { "k": "GDDR6-AiM 单颗规格", "v": "512 GFLOPS（BF16）、内部带宽 512 GB/s、外部带宽 32 GB/s", "date": "2024-08", "src": "https://hc2024.hotchips.org/assets/program/conference/day1/11_HC2024.SKhynix.GuhyunKim.rev920240822.pdf" }, { "k": "相对收益", "v": "特定计算速度最高 16 倍、功耗降低 80%（公司口径）", "date": "2022-02", "src": "https://news.skhynix.com/en/sk-hynix-develops-pim-next-generation-ai-accelerator/" } ], "product_line_of": "SK hynix", "public": true, "note": "同属一条线：CXL 近存计算模块 CMM-Ax（与 Marvell 合作，Arm 核 + CXL 控制器，2026-08 FMS 公布，OCP 2026-10 展出，https://biz.heraldcorp.com/article/10832627）、CuD（Compute-using-DRAM）。节点（1ynm 级，ISSCC 2022 论文）公司新闻稿未写，未单列。" }, { "name": "UPMEM", "layer": "n1", "url": "https://www.upmem.com", "hq": "格勒诺布尔", "region": "海外", "founded": 2015, "founded_month": "2015-01", "desc": "DDR4 DIMM 内置 DPU 处理核的商用 PIM 内存条", "where": "近存", "mode": "数字", "media": [ "DRAM" ], "stage": { "v": "已停", "date": "2025-07", "note": "2024-10 进入重整，2025-05 批准出售计划，2025-07 转清算", "src": "https://www.eiceu.com/report/upmem-eic-accelerator-processing-in-memory" }, "scene": [ "数据中心" ], "total_raised": "累计融资未见公司口径；EIC 报告列出 €2.50M 欧盟资助加 €4.1M 股权轮", "funding": [ { "date": "2023-09", "round": "股权轮", "amount": "€4.1M", "lead": "含 EIC Fund（领投方未标明）", "src": "https://www.eiceu.com/report/upmem-eic-accelerator-processing-in-memory" } ], "key_data": [ { "k": "DIMM 规格", "v": "每条 DIMM 128 个 DPU（350MHz），2,560 DPU 服务器", "date": "2025", "src": "https://theses.hal.science/tel-05398591/file/MOGNOL_Meven.pdf" } ], "owner": "2025-05 法院批准的出售计划中 Qualcomm France 为当事方，团队与 PIM 技术并入 Qualcomm；价格与条款未披露；2025-07 公司清算", "note": "各数据库记录的收购日期不一（2025-06 到 2025-09），以法院日期为准（EIC 报告引 Pappers/Societe.com）。第一轮草稿的 €7M 未找到一手出处，这里只写 EIC 报告的数字。" }, { "name": "XCENA", "layer": "n1", "url": "https://xcena.com/", "hq": "韩国（城市没查到）", "region": "海外", "founded": 2022, "founded_month": "2024", "desc": "MX1：CXL 接口的计算型内存控制器，在内存条内就近处理数据", "where": "近存", "mode": "数字", "media": [ "DRAM" ], "stage": { "v": "样片出货", "date": "2026-06", "note": "MX1 正与北美超大规模客户做验证，准备量产（电子新闻 2026-06-01）；2026-08 发布 MX1 量产产品线（仅聚合站转述）", "src": "https://www.etnews.com/20260601000245" }, "scene": [ "数据中心" ], "valuation": { "v": "$570M", "date": "2026-05", "kind": "B 轮投后估值，报道值", "src": "https://www.datacenterdynamics.com/en/news/south-korean-startup-xcena-raises-135m-for-its-mx1-memory-chip/" }, "total_raised": "累计融资约 $185M", "funding": [ { "date": "2026-05", "round": "B 轮", "amount": "$135M", "lead": "Altinum Investment、IMM Investment（共同领投；Corstone Asia、SBI Investment、Mirae Asset 跟投）", "post": "投后 $570M", "src": "https://www.datacenterdynamics.com/en/news/south-korean-startup-xcena-raises-135m-for-its-mx1-memory-chip/" } ], "customers": [ "Samsung", "Liqid" ], "note": "额外补充的一家。2022 年以 MetisX 创立，2024 年更名 XCENA（DCD）。fm 填的是更名年份，MX1 芯片首次公开时间没查到。customers 来自 Hot Chips 2026 的联合机架演示（XCENA 的 LinkedIn 帖，https://www.linkedin.com/posts/xcena_hotchips2026-cxl-aiinfrastructure-activity-7498873557196677120-xnT_，仅作线索）。TechCrunch 称芯片由 Samsung 代工，其余没核。" }, { "name": "东方算芯", "layer": "n1", "url": "https://www.dfsx.com/", "hq": "上海", "region": "中国", "founded": 2024, "founded_month": "2024-05", "desc": "软件定义加 3D 堆叠近存计算的大算力芯片 DF1000", "where": "近存", "mode": "数字", "media": [ "DRAM" ], "process_node": { "v": "14nm", "src": "https://www.chinaflashmarket.com/news/industry/184060" }, "stage": { "v": "流片", "date": "2026-07", "note": "DF1000 于 2026-07-13 发布；量产时间一手来源未查到", "src": "https://stcn.com/article/detail/4015764.html" }, "scene": [ "数据中心" ], "valuation": { "v": "122.75 亿元", "date": "2026-04", "kind": "A+ 轮投后估值，公司披露", "src": "https://www.cnstock.com/commonDetail/792418" }, "funding": [ { "date": "2026-04", "round": "A+ 轮", "amount": "未披露", "lead": "国家人工智能产业投资基金为股东（领投未标）", "post": "投后 122.75 亿元", "src": "https://www.cnstock.com/commonDetail/792418" } ], "key_data": [ { "k": "DF1000 算力", "v": "520 TFLOPS（BF16）", "date": "2026-07", "src": "https://stcn.com/article/detail/4015764.html" } ], "note": "技术源于清华大学集成电路学院，2026-09 完成股改，计划 2026Q4 启动 B 轮。存储介质未写明。「流片」是把「发布首颗芯片」归类，公司稿未写流片。" }, { "name": "紫光国芯 UniIC", "layer": "n1", "url": "https://www.unisemicon.com", "hq": "西安", "region": "中国", "founded": 2004, "desc": "SeDRAM：逻辑晶圆与 DRAM 晶圆混合键合 3D 堆叠，用于近存计算", "where": "近存", "mode": "数字", "media": [ "DRAM" ], "stage": { "v": "量产", "date": "2023-08", "note": "公司称嵌入式 DRAM（SeDRAM）技术已支持近 20 款产品研发与量产", "src": "https://www.unisemicon.com/index.php?m=content&c=index&a=show&catid=24&id=143" }, "scene": [ "数据中心", "边缘" ], "key_data": [ { "k": "访存带宽", "v": "可为算力芯片提供数十 TB/s 访存带宽，访存功耗较 HBM 方案减少 80% 以上（公司口径）", "date": "2025-11", "src": "https://www.unisemicon.com/index.php?m=content&c=index&a=show&catid=24&id=179" } ], "customers": [ "阿里达摩院" ], "note": "founded 取公司前身（2004 年英飞凌西安研发中心存储器事业部，证券时报）。阿里达摩院 ISSCC 2022 近存芯片基于其 SeDRAM 平台（出处为公众号，仅作线索，见达摩院条）。证券时报 2026-01-08 报道其开启上市辅导，此前 2024 年已登陆新三板（https://www.stcn.com/article/detail/3582125.html），未单独写 pub。主业是传统 DRAM 与模组，SeDRAM 是其中的近存方向。节点没查到。" }, { "name": "谦合益邦", "layer": "n1", "url": "https://www.agicmicro.com/", "hq": "北京", "region": "中国", "founded": 2024, "founded_month": "2026-08", "desc": "3D DRAM 与逻辑层垂直堆叠的存算一体芯片，用于云游戏等场景", "where": "近存", "mode": "数字", "media": [ "DRAM" ], "stage": { "v": "流片", "date": "2026-08", "note": "4 层 DRAM + 1 层逻辑的「4+1」芯片回片并点亮，进入工程落地阶段", "src": "https://ex.chinadaily.com.cn/exchange/partners/82/rss/channel/cn/columns/sz8srm/stories/WS6a880fcfa3105d3d7a27c494.html" }, "scene": [ "数据中心", "端侧" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2026-08", "round": "B 轮", "amount": "超 20 亿元", "lead": "领投方未标明（投资方含国调基金、中国移动链长基金等）", "src": "https://finance.eastmoney.com/a/202608263854419057.html" } ], "key_data": [ { "k": "公司口径提升", "v": "访存带宽较传统方案提升一个数量级，访存功耗与延时、单次处理成本各降一个数量级（未写对比基线）", "date": "2026-08", "src": "https://www.ithome.com/0/992/626.htm" } ], "customers": [ "网易有道", "中国移动链长基金" ], "note": "额外补充，可能与别组重叠（3D DRAM 存算）。网易孵化，网易有道为长期股东（zol 报道）。节点、带宽绝对值没查到。公司核心团队自 2018 年起探索，注册成立 2024-01（百度百科，仅作线索）。" }, { "name": "阿里达摩院 3D 堆叠 DRAM 存算芯片", "layer": "n1", "url": "http://news.21csp.com.cn/c15/202112/11410966.html", "hq": "杭州", "region": "中国", "founded": 2021, "founded_month": "2021-12", "desc": "混合键合 3D 堆叠 DRAM 加定制加速器，加速推荐系统的近存芯片", "where": "近存", "mode": "数字", "media": [ "DRAM" ], "stage": { "v": "研究", "date": "2022-02", "note": "研发成果，论文入选 ISSCC 2022；未见量产或商用报道", "src": "http://news.21csp.com.cn/c15/202112/11410966.html" }, "scene": [ "数据中心" ], "key_data": [ { "k": "推荐系统对比", "v": "相比 Xeon Gold 5200 平台，性能提升 10 倍以上、能效提升 300 倍以上", "date": "2021-12", "src": "http://news.21csp.com.cn/c15/202112/11410966.html" } ], "customers": [ "紫光国芯" ], "product_line_of": "阿里巴巴", "note": "核实结论：真实存在，是 2021-12 公布的研究芯片，存储层用异构集成嵌入式 DRAM（SeDRAM）。该芯片由紫光国芯完成实现，此事的出处只有公众号（mp.weixin.qq.com/s?__biz=MzIxODM3OTIxMg==&mid=2247486808），仅作线索。之后是否继续推进没查到；达摩院此芯片节点没查到。" }, { "name": "MemryX", "layer": "n2", "url": "https://memryx.ai", "hq": "安娜堡", "region": "海外", "founded": 2019, "desc": "Compute-at-memory 数据流边缘 AI 加速芯片 MX3", "where": "近存", "mode": "数字", "media": [ "SRAM" ], "stage": { "v": "量产", "date": "2025-03", "note": "B 轮用于扩大 MX3 量产并设计面向数据中心的 MX4", "src": "https://memryx.ai/news/memryx-raises-44m-series-b-to-scale-production-of-edge-ai-chips-and-expand-into-data-center-market/" }, "scene": [ "边缘", "端侧" ], "total_raised": "累计融资约 $63M（SDxCentral 口径）", "funding": [ { "date": "2025-03", "round": "B 轮", "amount": "$44M", "lead": "HarbourVest Partners 领投（NEOM Investment Fund、Arm IoT Fund、eLab Ventures、M Ventures、Motus Ventures 跟投）", "src": "https://memryx.ai/news/memryx-raises-44m-series-b-to-scale-production-of-edge-ai-chips-and-expand-into-data-center-market/" } ], "key_data": [ { "k": "MX3 算力", "v": "最高约 6 TFLOPS，单芯片平均功耗 0.6–2W；M.2 模块 24 TOPS", "date": "2025-03", "src": "https://www.sdxcentral.com/news/edge-ai-semiconductor-firm-memryx-raises-44m/" } ], "note": "存储介质、工艺节点没查到一手来源，media 为按 at-memory 描述的推定。" }, { "name": "Syntiant", "layer": "n2", "url": "https://www.syntiant.com", "hq": "尔湾", "region": "海外", "founded": 2017, "desc": "近存神经决策处理器 NDP，超低功耗端侧语音与传感推理", "where": "近存", "mode": "数字", "media": [ "其他" ], "stage": { "v": "量产", "date": "2026-07", "note": "NDP 已用于耳机、可穿戴、手机等；2026-07 提交 S-1 拟在纳斯达克上市（代码 SYTN）", "src": "https://thenextweb.com/news/syntiant-us-ipo-edge-ai-nasdaq" }, "scene": [ "端侧" ], "revenue": { "v": "$271.8M", "date": "2025", "kind": "S-1 披露全年总营收，绝大部分来自 2024 年底收购的传感器业务；其中 AI 业务约 $12.3M", "src": "https://content-archive.fast-edgar.com/20260706/A82VQ22EZZ2RO9ZU22ZS2ZYLAKNMZZ22Z28Q/" }, "also_in": [ ], "note": "pub 未填：截至 2026-10 查到仍未挂牌。AI 业务金额出自同花顺转载的 S-1 中文表格（http://news.10jqka.com.cn/field/sn/20260427/58803267.shtml），建议复核原文。存储介质、工艺节点、累计融资没查到一手来源（S-1 原文可查）。" }, { "name": "Untether AI", "layer": "n2", "url": "https://www.eetimes.com/untether-ai-shuts-down-engineering-team-joins-amd/", "hq": "多伦多", "region": "海外", "founded": 2018, "desc": "At-memory 架构 AI 推理芯片 speedAI，团队并入 AMD", "where": "近存", "mode": "数字", "media": [ "SRAM" ], "stage": { "v": "已停", "date": "2025-06", "note": "不再供应和支持 speedAI 与 imAIgine SDK", "src": "https://www.verdict.co.uk/amd-acquires-untether-ai-team/" }, "scene": [ "数据中心", "边缘" ], "total_raised": "累计融资未披露（第三方库约 $152M，未采）", "key_data": [ { "k": "speedAI240", "v": "1400 个定制 RISC-V 核，2 PFLOPS，20 TFLOPS/W，75W 卡", "date": "2025-06", "src": "https://m.techweb.com.cn/article/2025-06-06/2961699.shtml" } ], "owner": "2025-06 AMD 收购工程团队（技术与产品线不收购），公司停止供货，金额未披露", "note": "工艺节点没查到。" }, { "name": "Cerebras", "layer": "n3", "url": "https://www.cerebras.ai", "hq": "桑尼韦尔", "region": "海外", "desc": "晶圆级芯片，靠片上 SRAM 换带宽做推理与训练", "where": "相邻", "mode": "数字", "media": [ "SRAM" ], "process_node": { "v": "5nm（WSE-3）", "src": "https://www.forbes.com/sites/jonmarkman/2026/05/14/cerebras-priced-the-biggest-ipo-of-2026-into-a-market-huang-locked-up/" }, "stage": { "v": "量产", "date": "2026-05", "note": "2026-05-14 起在纳斯达克交易，代码 CBRS", "src": "https://thenextweb.com/news/cerebras-ipo-5-55-billion-biggest-tech-2026" }, "scene": [ "数据中心" ], "valuation": { "v": "$56.4B", "date": "2026-05", "kind": "IPO 定价 $185/股时的完全稀释估值", "src": "https://thenextweb.com/news/cerebras-ipo-5-55-billion-biggest-tech-2026" }, "revenue": { "v": "$510M", "date": "2025", "kind": "全年营收（招股文件口径，媒体转述）", "src": "https://thenextweb.com/news/cerebras-ipo-5-55-billion-biggest-tech-2026" }, "funding": [ { "date": "2026-02", "round": "H 轮", "amount": "$1B", "lead": "Tiger Global 领投（Benchmark、Fidelity、Atreides、Alpha Wave、Altimeter、AMD、Coatue 等参与）", "post": "投后约 $230 亿", "src": "https://www.cerebras.ai/press-release/cerebras-systems-raises-usd1-billion-series-h" } ], "key_data": [ { "k": "WSE-3 片上 SRAM", "v": "44GB，带宽 21 PB/s", "date": "2026-05", "src": "https://www.forbes.com/sites/jonmarkman/2026/05/14/cerebras-priced-the-biggest-ipo-of-2026-into-a-market-huang-locked-up/" } ], "customers": [ "OpenAI", "AWS", "G42" ], "public": true, "also_in": [ { "t": "Agent Infra 图谱 · 芯片推理", "u": "https://agent-infra-map.vercel.app/#g=r1" } ], "note": "IPO 募资约 $55.5 亿。OpenAI 为 2026-01 的多年 750MW 推理合同，AWS 见 Renaissance Capital 报道；G42 是 2024 年主要客户。成立年份没取到一手来源，未填。" }, { "name": "Graphcore", "layer": "n3", "url": "https://www.graphcore.ai", "hq": "布里斯托", "region": "海外", "founded": 2016, "desc": "每个核心带大块本地 SRAM 的 IPU，相邻代表", "where": "相邻", "mode": "数字", "media": [ "SRAM" ], "stage": { "v": "量产", "date": "2024-07", "note": "被软银收购后保留 Graphcore 名称，下一代 Izanagi 面向 SoftBank 的 Stargate，报道称 2026 年", "src": "https://www.jonpeddie.com/news/graphcores-ipu-doing-well-at-softbank/" }, "scene": [ "数据中心" ], "total_raised": "累计融资约 $684M（Jon Peddie Research 口径）", "owner": "2024-07-11 软银集团完成收购，成为全资子公司；JPR 称价格约 $5 亿，软银后续注资 $4.57 亿，价格双方未正式公布", "note": "未查到被收购后的营收和客户。" }, { "name": "Groq", "layer": "n3", "url": "https://groq.com", "hq": "山景城", "region": "海外", "founded": 2016, "desc": "片上大容量 SRAM 的 LPU 推理芯片，现转向 GroqCloud 云服务", "where": "相邻", "mode": "数字", "media": [ "SRAM" ], "stage": { "v": "量产", "date": "2026-08", "note": "GroqCloud 继续运营，并将在数据中心加入 Nvidia 系统；LPU 之后的新芯片计划未查到", "src": "https://www.businessinsider.com/nvidia-partners-groq-ai-services-groqcloud-2026-8" }, "scene": [ "数据中心" ], "valuation": { "v": "$3.5B", "date": "2026-08", "kind": "最新一轮投后估值，约为 2025-09 的 $6.9B 的一半（报道值）", "src": "https://techfundingnews.com/nvidia-backed-groq-raises-350m-at-3-5b-as-ai-inference-race-accelerates/" }, "total_raised": "累计融资未披露", "funding": [ { "date": "2026-08", "round": "未标轮次", "amount": "$350M", "lead": "Disruptive 领投，Nvidia 拟参与", "post": "投后 $3.5B", "src": "https://techfundingnews.com/nvidia-backed-groq-raises-350m-at-3-5b-as-ai-inference-race-accelerates/" }, { "date": "2026-06", "round": "未标轮次", "amount": "$650M", "lead": "现有股东参与（领投未查到）", "src": "https://www.pymnts.com/news/investment-tracker/2026/groq-seeks-650-million-amid-shift-to-ai-inference-neocloud-business/" } ], "owner": "2025-12-24 Nvidia 与 Groq 达成非独占技术授权并聘走创始人、总裁等，约 90% 员工转入 Nvidia（Axios 转述）；媒体称金额约 $200 亿，Nvidia 称并非收购；Groq 作为独立公司继续运营", "also_in": [ { "t": "Agent Infra 图谱 · 芯片推理", "u": "https://agent-infra-map.vercel.app/#g=r1" } ], "note": "owner 里的金额是投资方 / 媒体口径，双方未公布。无人收购，Groq 仍是私有独立公司。" }, { "name": "SambaNova", "layer": "n3", "url": "https://sambanova.ai", "hq": "圣何塞", "region": "海外", "desc": "数据流架构加大片上 SRAM 的推理芯片 SN50，相邻代表", "where": "相邻", "mode": "数字", "media": [ "SRAM", "HBM" ], "stage": { "v": "量产", "date": "2026-02", "note": "SN40L 在售；新一代 SN50 2026-02 发布，计划 2026 下半年出货", "src": "https://sambanova.ai/press/sambanova-unveils-fastest-chip-for-agentic-ai-collaborates-with-intel-and-raises-350m" }, "scene": [ "数据中心" ], "valuation": { "v": "$11B", "date": "2026-07", "kind": "F 轮首次交割投后估值（Dealroom / TechCrunch 报道值）", "src": "https://dealroom.co/news/137657-sambanova-banks-1b-series-f-first-close-at-11b-valuation/" }, "funding": [ { "date": "2026-07", "round": "F 轮（首次交割）", "amount": "$1B", "lead": "General Atlantic 领投", "post": "投后 $110 亿", "src": "https://dealroom.co/news/137657-sambanova-banks-1b-series-f-first-close-at-11b-valuation/" }, { "date": "2026-02", "round": "E 轮", "amount": "$350M", "lead": "公司稿未标领投（新投资方含 Vista Equity Partners、Intel Capital、Battery Ventures）", "src": "https://sambanova.ai/press/sambanova-unveils-fastest-chip-for-agentic-ai-collaborates-with-intel-and-raises-350m" } ], "customers": [ "SoftBank Corp." ], "also_in": [ { "t": "Agent Infra 图谱 · 芯片推理", "u": "https://agent-infra-map.vercel.app/#g=r1" } ], "note": "仍是独立私有公司。初稿线索称 Intel 收购谈判停滞，现状是 2026-02 与 Intel 做战略合作并获 Intel Capital 投资（sambanova.ai 新闻页标题称此前收购谈判据报告已告吹）。媒体称 SN50 2026 下半年出货，是否已出货未查到。" }, { "name": "CEA-Leti", "layer": "r1", "url": "https://www.leti-cea.fr/cea-tech/leti/Pages/actualites/Revue%20de%20Presse/Communique%20de%20presse/French-Team-led-by-CEA-Leti-Develops-First-Hybrid-Memory-Technology-Enabling-On-Chip-AI-Learning-and-Inference.aspx", "hq": "格勒诺布尔", "region": "海外", "founded": 2025, "founded_month": "2025-09", "desc": "铁电电容与忆阻器合一的混合存储，片上同时做训练与推理", "where": "存内", "mode": "模拟", "media": [ "RRAM", "其他" ], "stage": { "v": "研究", "date": "2025-09", "note": "Nature Electronics 论文公布的混合存储结构", "src": "https://www.electronicsweekly.com/news/research-news/ic-combines-machine-learning-and-analogue-inferencing-2025-09/" }, "scene": [ "边缘", "端侧" ], "key_data": [ { "k": "论文", "v": "A Ferroelectric-Memristor Memory for Both Training and Inference，Nature Electronics；可在两层金属间同一结构做铁电电容或忆阻器，片上训练精度有竞争力", "date": "2025-09", "src": "https://www.leti-cea.fr/cea-tech/leti/Pages/actualites/Revue%20de%20Presse/Communique%20de%20presse/French-Team-led-by-CEA-Leti-Develops-First-Hybrid-Memory-Technology-Enabling-On-Chip-AI-Learning-and-Inference.aspx" } ], "product_line_of": "CEA", "note": "资料偏少：没查到 TOPS/W 等量化指标，这条按器件/系统级研究写，节点未写。Leti 另有 RRAM 存内逻辑运算实验（2021 科学报告）和与 Weebit 的 RRAM 合作，未单列。" }, { "name": "IBM NorthPole", "layer": "r1", "url": "https://research.ibm.com/publications/neural-inference-at-the-frontier-of-energy-space-and-time", "hq": "阿蒙克", "region": "海外", "founded": 2023, "founded_month": "2023-10", "desc": "取消片外存储、计算与片上存储交织的神经推理研究芯片", "where": "近存", "mode": "数字", "process_node": { "v": "12nm", "src": "https://www.science.org/doi/10.1126/science.adh1174" }, "stage": { "v": "研究", "date": "2023-10", "note": "Science 论文公布的研究原型", "src": "https://www.science.org/doi/10.1126/science.adh1174" }, "scene": [ "边缘", "数据中心" ], "key_data": [ { "k": "ResNet-50 能效", "v": "相比同为 12nm 的 GPU，每瓦帧率高 25 倍；相比 H100（4nm）高 5 倍", "date": "2023-10", "src": "https://www.science.org/doi/10.1126/science.adh1174" }, { "k": "芯片规模", "v": "12nm、220 亿晶体管、800mm²、256 核", "date": "2023-10", "src": "https://www.science.org/doi/10.1126/science.adh1174" }, { "k": "时延", "v": "batch 1：106μs、9,454 FPS；batch 32：42,460 FPS", "date": "2023-10", "src": "https://www.science.org/doi/10.1126/science.adh1174" } ], "product_line_of": "IBM", "public": true, "note": "论文称从核内看是「存储靠近计算」，从外部看像「主动存储」，归近存。片上存储介质论文页摘要未写，没填 media。" }, { "name": "IBM 模拟 AI（Hermes PCM 芯片）", "layer": "r1", "url": "https://researchweb.draco.res.ibm.com/publications/a-64-core-mixed-signal-in-memory-compute-chip-based-on-phase-change-memory-for-deep-neural-network-inference", "hq": "苏黎世", "region": "海外", "founded": 2023, "founded_month": "2023-08", "desc": "14nm CMOS 集成相变存储器的 64 核模拟存内计算推理芯片", "where": "存内", "mode": "混合", "media": [ "PCM" ], "process_node": { "v": "14nm", "src": "https://researchweb.draco.res.ibm.com/publications/a-64-core-mixed-signal-in-memory-compute-chip-based-on-phase-change-memory-for-deep-neural-network-inference" }, "stage": { "v": "研究", "date": "2023-08", "note": "Nature Electronics 论文公布的研究芯片", "src": "https://www.eetimes.com/ibm-research-inference-chip-performance-results-released/" }, "scene": [ "边缘", "数据中心" ], "key_data": [ { "k": "峰值吞吐与能效", "v": "8 位 MVM 峰值 16.1–63.1 TOPS，能效 2.48–9.76 TOPS/W", "date": "2023-08", "src": "https://www.eetimes.com/ibm-research-inference-chip-performance-results-released/" }, { "k": "规模", "v": "64 个 AIMC 核，每核 256×256 交叉阵列，约 400 Gop/s/mm²", "date": "2023-08", "src": "https://www.electronicsweekly.com/news/research-news/ibms-analogue-ai-chip-achieves-400gop-s-mm2-2023-08/" } ], "product_line_of": "IBM", "public": true, "note": "论文：Nature Electronics 6:680–693，https://www.nature.com/articles/s41928-023-01010-1。另有同年 34 块 PCM 阵列、3,500 万 PCM 器件的语音识别芯片（Nature Electronics 2023-09，来源 https://stm.las.ac.cn 转述，未单列）。第一轮草稿提到的 2026-01 3D 模拟架构报道这次没复核。" }, { "name": "imec", "layer": "r1", "url": "https://gf.com/de/gf-press-release/imec-and-globalfoundries-announce-breakthrough-ai-chip-bringing-deep-neural-network/", "hq": "鲁汶", "region": "海外", "founded": 2020, "founded_month": "2020-07", "desc": "与 GlobalFoundries 合作的 SRAM 内模拟计算推理加速器 AnIA", "where": "存内", "mode": "模拟", "media": [ "SRAM" ], "process_node": { "v": "22nm FD-SOI", "src": "https://gf.com/de/gf-press-release/imec-and-globalfoundries-announce-breakthrough-ai-chip-bringing-deep-neural-network/" }, "stage": { "v": "研究", "date": "2020-07", "note": "硬件演示；称 AiMC 将作为 22FDX 平台的特性提供给客户", "src": "https://gf.com/de/gf-press-release/imec-and-globalfoundries-announce-breakthrough-ai-chip-bringing-deep-neural-network/" }, "scene": [ "端侧", "边缘" ], "key_data": [ { "k": "峰值能效", "v": "最高 2,900 TOPS/W（公司称记录水平）", "date": "2020-07", "src": "https://www.electronicsweekly.com/news/business/ai-2-2020-07/" } ], "customers": [ "GlobalFoundries" ], "product_line_of": "imec", "note": "founded 填的是 AnIA 首次公开年份，不是 imec 成立年份。2,900 TOPS/W 是峰值、低精度条件，新闻稿未写精度，对比时需标注。imec 之后的存内工作（如 IMAGINE 加速器，22nm FD-SOI，TU Delft 库论文）这次没单列。" }, { "name": "Intel 存内计算研究（18A 数字存内）", "layer": "r1", "url": "https://ieeexplore.ieee.org/document/11409207", "hq": "圣克拉拉", "region": "海外", "founded": 2026, "founded_month": "2026-02", "desc": "Intel 18A 工艺全可综合数字存内计算加速器测试芯片（ISSCC 2026）", "where": "存内", "mode": "数字", "media": [ "其他" ], "process_node": { "v": "Intel 18A", "src": "https://ieeexplore.ieee.org/document/11409207" }, "stage": { "v": "研究", "date": "2026-02", "note": "ISSCC 2026 论文 30.9 的测试芯片", "src": "https://ieeexplore.ieee.org/document/11409207" }, "scene": [ "端侧", "边缘" ], "key_data": [ { "k": "能效", "v": "0.4V、25% 输入活跃度下峰值 147 TOPS/W，功耗 6.1 mW", "date": "2026-02", "src": "https://ieeexplore.ieee.org/document/11409207" }, { "k": "面积效率与算力", "v": "250 TOPS/mm²，21.5 TOPS，2.62 GHz（1.1V）", "date": "2026-02", "src": "https://ieeexplore.ieee.org/document/11409207" } ], "product_line_of": "Intel", "public": true, "note": "存储单元为 12T 锁存器（非标准 SRAM），故 media 填「其他」。同期 Intel 在 ISSCC 2026 另发 10.6「混合键合 56 核 DNN 处理器」（ASCII.jp 转述，https://ascii.jp/elem/000/004/384/4384110/），2022 VLSI 有「Compute Near LLC」近存实现，均未单列。Loihi 归类脑组。" }, { "name": "Samsung MRAM 存内计算", "layer": "r1", "url": "https://www.nature.com/articles/s41586-021-04196-6", "hq": "水原", "region": "海外", "founded": 2022, "founded_month": "2022-01", "desc": "64×64 MRAM 交叉阵列，用「电阻求和」做模拟乘加，Nature 论文", "where": "存内", "mode": "模拟", "media": [ "MRAM" ], "process_node": { "v": "28nm", "src": "https://semiengineering.com/a-crossbar-array-of-magnetoresistive-memory-devices-for-in-memory-computing/" }, "stage": { "v": "研究", "date": "2022-01", "note": "Nature 论文展示的研究阵列", "src": "https://www.eenewseurope.com/en/samsung-shows-mram-can-support-in-memory-computing/" }, "scene": [ "端侧" ], "key_data": [ { "k": "两层感知机 MNIST", "v": "识别准确率 93.23%（软件基线 95.24%）", "date": "2022-01", "src": "https://www.eenewseurope.com/en/samsung-shows-mram-can-support-in-memory-computing/" }, { "k": "VGG-8 仿真", "v": "带实测误差的仿真准确率 98.86%（软件基线 99.28%）", "date": "2022-01", "src": "https://www.eenewseurope.com/en/samsung-shows-mram-can-support-in-memory-computing/" }, { "k": "人脸检测", "v": "单层实现，准确率 93.4%", "date": "2022-01", "src": "https://www.ixbt.com/news/2022/01/13/samsung-udalos-to-chto-ne-udavalos-nikomu-kompanija-pervoj-v-mire-realizovala-process-vychislenij-v-operativnoj-pamjati.html" } ], "product_line_of": "Samsung", "public": true, "note": "论文：A crossbar array of magnetoresistive memory devices for in-memory computing，Nature 2022，出处 Samsung 新闻稿 https://news.samsungsemiconductor.com/kr/ 。能效 TOPS/W 数字这次没拿到一手原文，未写。第一轮草稿提到的 MRAM 推进到 5nm 的报道没复核。" }, { "name": "Stanford/UCSD NeuRRAM", "layer": "r1", "url": "https://www.nature.com/articles/s41586-022-04992-8", "hq": "斯坦福", "region": "海外", "founded": 2022, "founded_month": "2022-08", "desc": "48 核 RRAM 存内计算芯片，可重构支持多种模型，Nature 论文", "where": "存内", "mode": "模拟", "media": [ "RRAM" ], "stage": { "v": "研究", "date": "2022-08", "note": "Nature 论文展示的芯片", "src": "https://par.nsf.gov/biblio/10406027-compute-memory-chip-based-resistive-random-access-memory" }, "scene": [ "端侧" ], "key_data": [ { "k": "能效", "v": "各计算位宽下约为此前最优 RRAM-CIM 芯片的 2 倍", "date": "2022-08", "src": "https://par.nsf.gov/biblio/10406027-compute-memory-chip-based-resistive-random-access-memory" }, { "k": "任务准确率", "v": "MNIST 99.0%、CIFAR-10 85.7%、语音命令 84.7%", "date": "2022-08", "src": "https://par.nsf.gov/biblio/10406027-compute-memory-chip-based-resistive-random-access-memory" } ], "product_line_of": "Stanford / UC San Diego", "note": "Stanford 与 UCSD 联合，48 核（eeNews 转述，https://ceni.ucsd.edu/press/archive?year=2022）。工艺节点、TOPS/W 绝对值这次没拿到一手原文，未写。" }, { "name": "TSMC SRAM CIM", "layer": "r1", "url": "https://research.tsmc.com/english/collaborations/events/event/conferences-5.html", "hq": "新竹", "region": "海外", "founded": 2021, "founded_month": "2021-02", "desc": "ISSCC 论文线：全数字及模拟 SRAM 存内计算宏，另有 ReRAM、PCM 版", "where": "存内", "mode": "数字", "media": [ "SRAM", "RRAM", "PCM" ], "process_node": { "v": "5nm（数字 CIM 宏，ISSCC 2022）", "src": "https://ieeexplore.ieee.org/ielx7/9731529/9731102/09731761.pdf" }, "stage": { "v": "研究", "date": "2022-02", "note": "ISSCC 论文发表的测试宏，非产品", "src": "https://ieeexplore.ieee.org/ielx7/9731529/9731102/09731761.pdf" }, "scene": [ "端侧", "边缘" ], "key_data": [ { "k": "5nm 全数字 CIM 宏", "v": "254 TOPS/W、221 TOPS/mm²（4b×4b，ISSCC 2022 11.6）", "date": "2022-02", "src": "https://ieeexplore.ieee.org/ielx7/9731529/9731102/09731761.pdf" }, { "k": "22nm ReRAM CIM 宏", "v": "11.91 至 195.7 TOPS/W（8b MAC）", "date": "2021", "src": "https://research.tsmc.com/english/collaborations/events/event/conferences-5.html" }, { "k": "40nm PCM CIM 宏", "v": "20.5–65.0 TOPS/W（8b）", "date": "2022-02", "src": "https://ieeexplore.ieee.org/ielx7/9731529/9731102/09731761.pdf" } ], "product_line_of": "TSMC", "public": true, "note": "founded 取 TSMC 22nm 全数字 CIM（ISSCC 2021 16.4，89 TOPS/W）的发表年份，该篇只在他人论文参考文献里见到，未拿到原页，故 data 里未写。ISSCC 论文线，部分与台湾高校合作，不是商品。" }, { "name": "中科院微电子所 3D RRAM 存内计算宏", "layer": "r1", "url": "https://www.ime.cas.cn/zhxx/zhxw/202207/t20220729_6494990.html", "hq": "北京", "region": "中国", "founded": 2022, "founded_month": "2022-07", "desc": "三维垂直结构阻变存储器的存内计算宏芯片，Nature Electronics", "where": "存内", "mode": "模拟", "media": [ "RRAM" ], "stage": { "v": "研究", "date": "2022-07", "note": "Nature Electronics 论文公布的宏单元芯片", "src": "https://www.ime.cas.cn/zhxx/zhxw/202207/t20220729_6494990.html" }, "scene": [ "边缘" ], "key_data": [ { "k": "测试配置", "v": "输入 8 位、权重 9 位、输出 22 位（多值自选通 3D 垂直 RRAM 阵列）", "date": "2022-07", "src": "https://www.ime.cas.cn/zhxx/zhxw/202207/t20220729_6494990.html" }, { "k": "论文", "v": "A computing-in-memory macro with three-dimensional resistive random-access memory，Nature Electronics 5:469–477", "date": "2022-07", "src": "https://www.ime.cas.cn/zhxx/zhxw/202207/t20220729_6494990.html" } ], "product_line_of": "中国科学院微电子研究所", "note": "资料偏少：所页摘要里位密度、能效数字被截断，没拿到完整指标，不写。节点没查到。该所页面（ime.cas.cn）抓取时 404，数据来自搜索结果摘要。" }, { "name": "北京大学 RRAM 模拟矩阵计算芯片", "layer": "r1", "url": "http://xinhuanet.com.cn/auto/20251017/843e5e268856424f98d88871b5d4ce46/c.html", "hq": "北京", "region": "中国", "founded": 2025, "founded_month": "2025-10", "desc": "基于阻变存储器的高精度可扩展模拟矩阵方程求解芯片", "where": "存内", "mode": "模拟", "media": [ "RRAM" ], "stage": { "v": "研究", "date": "2025-10", "note": "Nature Electronics 论文公布，2025-10-13 发表", "src": "http://www.jingjiribao.cn/static/detail.jsp?id=616397" }, "scene": [ "边缘", "数据中心" ], "key_data": [ { "k": "计算精度", "v": "16×16 矩阵方程求解达 24 位定点精度，10 次迭代相对误差可低至 10⁻⁷ 量级", "date": "2025-10", "src": "http://www.jingjiribao.cn/static/detail.jsp?id=616397" }, { "k": "对比数字处理器", "v": "128×128 规模下吞吐量可达顶级数字处理器的 1000 倍以上，同精度能效高 100 倍以上（团队口径）", "date": "2025-10", "src": "http://xinhuanet.com.cn/auto/20251017/843e5e268856424f98d88871b5d4ce46/c.html" } ], "product_line_of": "北京大学", "note": "论文：Precise and scalable analogue matrix equation solving using resistive random-access memory chips（Nature Electronics，题名见澎湃 https://www.thepaper.cn/newsDetail_forward_31791811）。做的是矩阵求逆而非神经网络推理，与其他存内芯片不是同一类任务。节点没查到。" }, { "name": "清华大学 忆阻器存算芯片", "layer": "r1", "url": "https://www.tsinghua.edu.cn/info/3045/110193.htm", "hq": "北京", "region": "中国", "founded": 2023, "founded_month": "2023-10", "desc": "全系统集成、支持片上学习的忆阻器存算一体芯片，发表于 Science", "where": "存内", "mode": "混合", "media": [ "RRAM" ], "stage": { "v": "研究", "date": "2023-10", "note": "Science 论文公布的原型芯片", "src": "https://www.cas.cn/zt/kjzt/2023kjpd/2023kjrb1/202312/t20231229_4993289.shtml" }, "scene": [ "边缘", "端侧" ], "key_data": [ { "k": "片上学习能耗", "v": "相同任务下仅为先进工艺 ASIC 系统的约 3%，能效约 75 倍", "date": "2023-10", "src": "https://www.ncsti.gov.cn/kjdt/xwjj/202310/t20231011_134883.html" }, { "k": "论文", "v": "Science 2023, 381: 1205–1211", "date": "2023-10", "src": "https://www.ee.tsinghua.edu.cn/info/1076/4366.htm" } ], "product_line_of": "清华大学", "note": "节点没查到。ISSCC 2026 清华、华为、字节联合的 28nm 混合存内芯片（QPS 提升 66 倍、QPS/W 提升 181 倍）只有 IC 资料网转述（https://www.icpdf.com/design/3022.html），未单列。" }, { "name": "4DS Memory", "layer": "u1", "url": "https://www.4dsmemory.com", "hq": "弗里蒙特", "region": "海外", "founded": 2007, "desc": "界面开关型 PCMO ReRAM，面向 AI 处理器的高带宽持久存储", "where": "上游", "media": [ "RRAM" ], "stage": { "v": "研究", "date": "2025-04", "note": "imec 完成第六批平台晶圆并寄出，初步结果预计 2025 Q2", "src": "https://www.4dsmemory.com/wp-content/uploads/2025/04/61261563.pdf" }, "total_raised": "上市公司（ASX: 4DS），累计融资未汇总", "customers": [ "imec（开发协议）", "Infineon（设计协议）", "HGST / Western Digital（联合开发协议）" ], "public": true, "note": "额外补充；2026-07-31 有公司更新，仅见转载页（hotcopper），内容未核；规模很小，工艺节点没查到。" }, { "name": "Avalanche Technology", "layer": "u1", "url": "https://www.avalanche-technology.com", "hq": "弗里蒙特", "region": "海外", "desc": "航天/国防级 STT-MRAM，22nm 工艺量产，非存算", "where": "上游", "media": [ "MRAM" ], "process_node": { "v": "22nm bulk CMOS", "src": "https://www.avalanche-technology.com/wp-content/uploads/Space-Computing-Conference-2025-Web.pdf" }, "stage": { "v": "量产", "date": "2025", "note": "公司称自 2019 年起向工业客户批量出货；2026-03 完成美国政府合同第一阶段（缩小 MTJ，目标密度提升 16 倍）", "src": "https://www.advfn.com/stock-market/stock-news/97952216/avalanche-technology-achieves-phase-one-milestone" }, "scene": [ "边缘" ], "total_raised": "累计融资未披露", "customers": [ "Rincon Research", "Orbion Space Technology", "NHanced Semiconductors（2.5D 集成合作）" ], "note": "与存算无直接产品；作 MRAM 器件上游代表，是否保留可再议；成立年份与融资没查到一手。" }, { "name": "Crossbar Inc.", "layer": "u1", "url": "https://crossbar-inc.com", "hq": "圣克拉拉", "region": "海外", "founded": 2010, "desc": "ReRAM 器件/IP 公司，现以 22nm ReRAM 做安全处理芯片 Daric", "where": "上游", "media": [ "RRAM" ], "process_node": { "v": "22nm", "src": "https://crossbar-inc.com/" }, "total_raised": "累计融资未找到一手出处（CB Insights 页显示约 $120M，仅作线索未采）", "note": "主业已转向加密资产安全芯片（Daric SPU），与存算关联弱，建议考虑是否保留；ReRAM 早年面向存算与存储，官网当前未强调存算；量产状态与估值没查到。" }, { "name": "Everspin", "layer": "u1", "url": "https://www.everspin.com", "hq": "钱德勒（美国亚利桑那）", "region": "海外", "desc": "独立 MRAM 供应商，Toggle 与 STT-MRAM 产品", "where": "上游", "media": [ "MRAM" ], "stage": { "v": "量产", "date": "2026-09", "note": "MRAM 产品季度销售 $16.9M", "src": "https://stockanalysis.com/stocks/mram/transcripts/725395-q3-2026/" }, "scene": [ "边缘", "数据中心" ], "revenue": { "v": "$20.6M，同比增 47%", "date": "2026 Q3", "kind": "季度总营收，公司公布", "src": "https://investor.wedbush.com/wedbush/article/bizwire-2026-10-28-everspin-reports-unaudited-third-quarter-2026-financial-results" }, "public": true, "note": "NASDAQ: MRAM；Q3 GAAP 净亏约 $7.2M，含诉讼费用 $6M（报道）；与存算的关联是作存储器件/IP，本身不做存算；founded 没核。" }, { "name": "Ferroelectric Memory Co (FMC)", "layer": "u1", "url": "https://www.ferroelectric-memory.com", "hq": "德累斯顿", "region": "海外", "founded": 2016, "desc": "铁电氧化铪存储 DRAM+ 与 3D CACHE+，面向 AI 数据中心", "where": "上游", "media": [ "FeFET" ], "stage": { "v": "研究", "date": "2026-01", "note": "heise 报道：FMC 目前尚无铁电存储芯片产品，技术只能作嵌入式模块集成，与 GlobalFoundries 合作", "src": "https://www.heise.de/en/news/Memory-chip-company-FMC-keeps-working-on-semiconductor-plant-near-Magdeburg-11151026.html" }, "scene": [ "数据中心", "边缘" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2025-11", "round": "C 轮（含公共资金）", "amount": "€100M（股权 €77M + 公共资金 €23M）", "lead": "HV Capital、DTCF 共同领投（Vsquared Ventures）", "src": "https://www.linkedin.com/feed/update/urn:li:activity:7394706314775588864" } ], "note": "计划在萨克森-安哈尔特建厂，取决于融资（heise）；领投方信息来自 HV Capital 官方领英帖；FeFET/FeCAP 构件；估值没查到。" }, { "name": "Intrinsic Semiconductor Technologies", "layer": "u1", "url": "https://www.intrinsicsemi.com", "hq": "伦敦", "region": "海外", "founded": 2017, "desc": "硅氧化物 RRAM，面向 MCU/边缘 AI 的嵌入式非易失存储，UCL 分拆", "where": "上游", "media": [ "RRAM" ], "scene": [ "边缘", "端侧" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2023-03", "round": "股权融资", "amount": "£7M（另有 £1M Innovate UK 拨款）", "lead": "Octopus Ventures（IP Group、UCL Technology Fund 跟投）", "src": "https://semiiphub.com/pulse/news/intrinsic-semiconductor-rram-funding" } ], "note": "公司称读速快 10 至 100 倍、写速快 1000 倍（对比现有方案，口径未明）；技术方向包括模拟存储与类脑器件；PitchBook 显示 2025-02 另有一轮约 $8.1M，仅聚合页，未采；量产状态没查到。" }, { "name": "Numem", "layer": "u1", "url": "https://www.numem.com", "hq": "桑尼韦尔", "region": "海外", "desc": "STT-MRAM（NuRAM）加 SmartMem 存内计算子系统 IP 与 MRAM 芯粒", "where": "上游", "media": [ "MRAM" ], "scene": [ "边缘", "数据中心" ], "total_raised": "累计融资未披露（PitchBook 显示 A 轮约 $9.3M，聚合页未采）", "funding": [ { "date": "2023-09", "round": "A 轮", "amount": "未披露", "lead": "Cambium Capital、Doorga Capital", "src": "https://www.numem.com/wp-content/uploads/2024/03/Series-A-Funding-PR-090123.pdf" } ], "customers": [ "IC'ALPS（联合开发含 MRAM 的 AI 加速 SoC）" ], "note": "MRAM 芯粒原计划 2025 年底送样（Electronics Weekly 2025-01 报道），实际送样没查到；成立年份各库口径不一（2016/2018），省略；联合开发见 https://www.mram-info.com/numem 。" }, { "name": "Weebit Nano", "layer": "u1", "url": "https://www.weebit-nano.com", "hq": "霍德哈沙隆（以色列，ASX 上市）", "region": "海外", "desc": "嵌入式 ReRAM IP 授权；入选韩国国家模拟存内计算项目", "where": "上游", "media": [ "RRAM" ], "stage": { "v": "流片", "date": "2026-07", "note": "3 家客户已用 Weebit ReRAM 流片，其中 1 家已跑通软件的功能样片", "src": "https://www.weebit-nano.com/news/press-releases/weebit-nano-expands-licensing-agreements-with-key-customers-three-customer-chip-designs-taped-out-to-date/" }, "scene": [ "端侧", "边缘" ], "revenue": { "v": "A$15.276M", "date": "FY2026（截至 2026-06-30）", "kind": "年度营业收入，Appendix 4E", "src": "https://kalkinemedia.com/au/news/announcements/weebit-nano-reports-153-million-revenue-for-fiscal-year-2026" }, "total_raised": "上市公司，累计融资未汇总", "funding": [ { "date": "2026-07", "round": "配售 + 股东 SPP", "amount": "合计 A$102M（SPP A$15M）", "lead": "机构配售", "src": "https://www.weebit-nano.com/news/press-releases/weebit-nano-expands-licensing-agreements-with-key-customers-three-customer-chip-designs-taped-out-to-date/" } ], "key_data": [ { "k": "韩国模拟存内计算项目目标能效", "v": "约 200 TOPS/W（项目目标，非实测）", "date": "2026-03", "src": "https://semiiphub.com/pulse/news/weebit-nano-reram-korean-national-compute-in-memory-program" } ], "customers": [ "Texas Instruments", "onsemi", "DB HiTek", "AnalogAI（韩国项目参与方）" ], "public": true, "note": "ASX: WBT；公司季报 2026-07 写 FY26 营收指引上调至至少 A$13.5M，而 Appendix 4E 报 A$15.276M，两处口径以后者为准；韩国项目 2026-03-05 公告：ReRAM 为交叉阵列模拟存内的基础器件，DB HiTek 制造，参与方含 ETRI、首尔大学等；成立年份未核。" }, { "name": "Applied Materials", "layer": "u2", "url": "https://www.appliedmaterials.com", "hq": "圣克拉拉", "region": "海外", "founded": 1967, "desc": "设备龙头；公开过 b-ReRAM 模拟存内 SoC 原型及 MRAM/ReRAM 沉积设备", "where": "上游", "mode": "模拟", "media": [ "RRAM" ], "process_node": { "v": "65nm（b-ReRAM 存内计算 SoC 原型）", "src": "https://www.booksci.cn/literature/129779494.htm" }, "stage": { "v": "研究", "date": "2022-12", "note": "论文展示 BEOL 集成 b-ReRAM 的 65nm CMOS SoC 原型，关键词识别准确率与软件量化模型相当", "src": "https://www.booksci.cn/literature/129779494.htm" }, "key_data": [ { "k": "b-ReRAM 能效", "v": "98.5 TOPS/W（关键词识别）；128 个电导等级，25k 器件分布于 15 个 die", "date": "2022-12", "src": "https://www.booksci.cn/literature/129779494.htm" } ], "customers": [ "Arm、Symetrix（2018 DARPA ERI CeRAM 神经形态开关项目合作方）" ], "public": true, "note": "设备线另有 2019 年 Endura Impulse PVD（PCRAM/ReRAM）与 Clover MRAM PVD 平台公告（https://investors.appliedmaterials.com/node/22826/pdf ）；DARPA 项目见 https://www.appliedmaterials.com/tw/zh_tw/news-and-media/press-releases/applied-materials-selected-by-darpa-tw.html ；论文页为学术聚合页，会议名未在页内核实；只算研究线加设备，本身非存算厂商。" }, { "name": "Arago", "layer": "x1", "url": "https://www.arago.inc/", "hq": "巴黎", "region": "海外", "founded": 2024, "desc": "模拟/数字电子加硅光加自由空间光学的推理处理器 JEF", "where": "相邻", "mode": "光", "media": [ "光" ], "stage": { "v": "研究", "date": "2025-07", "note": "公司称早期结果显示可跑标准框架的 AI 模型，尚无量产信息", "src": "https://www.datacenterdynamics.com/en/news/photonic-chip-startup-arago-closes-26m-seed-round/" }, "scene": [ "数据中心" ], "total_raised": "累计融资约 $26M（仅种子轮）", "funding": [ { "date": "2025-07", "round": "种子轮", "amount": "$26M", "lead": "Earlybird、Protagonist、Visionaries Tomorrow 共同领投", "src": "https://www.datacenterdynamics.com/en/news/photonic-chip-startup-arago-closes-26m-seed-round/" } ], "note": "额外补充的第 7 家光子代表；官网域名没查到，url 用 EE Times 报道页；称能耗降低 10 倍为公司口径。" }, { "name": "Celestial AI", "layer": "x1", "url": "https://www.celestial.ai", "hq": "圣克拉拉", "region": "海外", "founded": 2020, "desc": "Photonic Fabric 光互连，解耦算力与内存；光互连而非光计算，作相邻代表", "where": "相邻", "mode": "光", "media": [ "光" ], "scene": [ "数据中心" ], "valuation": { "v": "$2.5B", "date": "2025-03", "kind": "C1 轮投后估值，报道值", "src": "https://www.nasdaq.com/articles/chip-startup-celestial-ai-hits-25b-valuation" }, "total_raised": "累计融资超 $515M（公司口径，Nasdaq 转述）", "funding": [ { "date": "2025-03", "round": "C1 轮", "amount": "$250M", "lead": "Fidelity Management & Research（BlackRock、Tiger Global、Maverick 等跟投）", "post": "投后 $2.5B", "src": "https://www.nasdaq.com/articles/chip-startup-celestial-ai-hits-25b-valuation" } ], "key_data": [ { "k": "互连能效与时延（Marvell 引述）", "v": "延迟低于 150ns，约 2.5 pJ/bit", "date": "2026-02", "src": "https://www.nasdaq.com/articles/how-mrvls-celestial-ai-acquisition-fills-its-critical-technology-gap" } ], "owner": "2026-02-02 被 Marvell 收购完成：首期约 $32.5 亿（现金 $10 亿 + 约 2,720 万股），按营收里程碑最高再付约 $22.5 亿，合计最高约 $55 亿", "note": "收购公告与完成公告见 Marvell 投资者页：https://investor.marvell.com/sec-filings/all-sec-filings/content/0001193125-26-032861/d45933dex991.htm ；按任务要求写 owner；Marvell 预计其业务 FY2028 Q4 年化收入达 $5 亿。" }, { "name": "Lightmatter", "layer": "x1", "url": "https://lightmatter.co", "hq": "山景城", "region": "海外", "founded": 2017, "desc": "硅光子公司，主推 3D 光互连 Passage，另有光计算处理器 Envise", "where": "相邻", "mode": "光", "media": [ "光" ], "scene": [ "数据中心" ], "valuation": { "v": "$4.4B", "date": "2024-10", "kind": "投后估值，公司公布", "src": "https://www.businesswire.com/news/home/20241016498931/en/Lightmatter-Raises-$400M-Series-D-Quadruples-Valuation-to-$4.4B-as-Photonics-Leader-for-Next-Gen-AI-Data-Centers" }, "total_raised": "累计融资约 $850M（公司口径）", "funding": [ { "date": "2024-10", "round": "D 轮", "amount": "$400M", "lead": "T. Rowe Price 旗下基金（Fidelity、GV 跟投）", "post": "投后 $4.4B", "src": "https://www.businesswire.com/news/home/20241016498931/en/Lightmatter-Raises-$400M-Series-D-Quadruples-Valuation-to-$4.4B-as-Photonics-Leader-for-Next-Gen-AI-Data-Centers" } ], "note": "现阶段商业重心是光互连（Passage），光计算只是相邻代表；Series D 之后未查到更新的融资；官网域名凭记忆未核。公司未点名客户（Reuters 2024-10 报道）。" }, { "name": "Lumai", "layer": "x1", "url": "https://lumai.ai", "hq": "牛津", "region": "海外", "founded": 2022, "desc": "自由空间 3D 光学张量引擎，混合数字控制，做大模型推理服务器", "where": "相邻", "mode": "光", "media": [ "光" ], "stage": { "v": "样片出货", "date": "2026-04", "note": "首款 Iris Nova 推理服务器开放给超大规模云、新云等评估，运行 Llama 8B/70B；评估系统而非量产", "src": "https://www.photonics.com/Articles/Lumai-Launches-Optical-Computing-System-for/p6/vo273/i1993/a72191" }, "scene": [ "数据中心" ], "total_raised": "累计融资未披露（第三方库约 $13.8M，未采）", "funding": [ { "date": "2025-04", "round": "A 轮", "amount": "超过 $10M", "lead": "Constructor Capital（IP Group、PhotonVentures、Journey Ventures、LIFTT、Qubits Ventures、State Farm Ventures、TIS 跟投）", "src": "https://www.enterpriseaiworld.com/Articles/News/News/Lumai-Raises-More-Than-$10M-in-Funds-Brings-its-Optical-Processing-Breakthrough-Forward-to-Address-Limitations-of-AI-Compute-168838.aspx" } ], "note": "牛津大学分拆；官网写 2022 年成立；估值没查到；公司宣称性能 50 倍、功耗 10% 均为公司口径。" }, { "name": "Neurophos", "layer": "x1", "url": "https://www.neurophos.com", "hq": "奥斯汀", "region": "海外", "desc": "超表面光调制器做光学处理单元（OPU），瞄准数据中心推理", "where": "相邻", "mode": "光", "media": [ "光" ], "scene": [ "数据中心" ], "total_raised": "累计融资 $118M（公司口径）", "funding": [ { "date": "2026-01", "round": "A 轮", "amount": "$110M", "lead": "Gates Frontier（M12、Carbon Direct Capital、Aramco Ventures、Bosch Ventures 等跟投）", "src": "https://www.neurophos.com/110m-raise" } ], "key_data": [ { "k": "芯片规模与性能（公司宣称）", "v": "单芯片 100 万余个微米级光处理单元，称性能与能效最高达当前领先芯片 100 倍", "date": "2026-01", "src": "https://www.neurophos.com/110m-raise" } ], "note": "仍在研发，称融资用于推进制造开发（Network World 2026-01 报道）；成立年份、工艺节点、客户没查到一手出处。" }, { "name": "Q.ANT", "layer": "x1", "url": "https://qant.com", "hq": "斯图加特", "region": "海外", "founded": 2018, "desc": "薄膜铌酸锂（TFLN）光子处理器，用光做 AI 矩阵与非线性运算", "where": "相邻", "mode": "光", "media": [ "光" ], "stage": { "v": "样片出货", "date": "2025-11", "note": "发布第二代 NPU 2 及可训练的软件栈；样片出货口径为推断，原文只写发布", "src": "https://www.eetimes.com/q-ant-raises-series-a-debuts-second-gen-tfln-photonic-chip/" }, "scene": [ "数据中心" ], "total_raised": "累计融资约 $80M（EE Times 2025-11 报道）", "funding": [ { "date": "2025-07", "round": "A 轮", "amount": "€62M", "lead": "Cherry Ventures、UVC Partners、imec.xpand 共同领投", "src": "https://qant.com/press-releases/qant-raises-62-million-euro-to-transform-the-future-of-computing-with-photonic-processing/" } ], "key_data": [ { "k": "算力", "v": "NPU 2 约 8 GOPS", "date": "2025-11", "src": "http://www.iccsz.com//Site/CN/News/2025/11/27/20251127011350250488.htm" } ], "customers": [ "Jülich Supercomputing Centre（4 年合作）" ], "note": "CEO 在 EE Times 表态不主攻大模型推理，瞄准图像、音频、视频网络；2018 年自 TRUMPF 分拆（Q.ANT 德文新闻稿）；量产状态没查到。" }, { "name": "曦智科技 Lightelligence", "layer": "x1", "url": "https://www.lightelligence.ai", "hq": "上海", "region": "中国", "founded": 2017, "desc": "光电混合算力：片上/片间光网络加光子矩阵计算（oMAC）", "where": "相邻", "mode": "光", "media": [ "光" ], "scene": [ "数据中心" ], "valuation": { "v": "总市值超 814 亿港元", "date": "2026-04-28", "kind": "上市首日收盘市值，媒体报道", "src": "https://m.cnfin.com/yw-lb//zixun/20260428/4406000_1.html" }, "revenue": { "v": "1.06 亿元（光互连 8428 万元，光计算 2020 万元）", "date": "2025", "kind": "2025 年营业收入，招股书披露，媒体转述", "src": "http://www.iccsz.com/site/cn/News/2026/04/28/20260428101610657109.htm" }, "funding": [ { "date": "2025-04", "round": "C4 轮", "amount": "3 亿元", "lead": "未写明", "post": "投后估值 78 亿元", "src": "https://www.21jingji.com/article/20260429/bf59b113eb092d8004d33997048b7aeb.html" }, { "date": "2026-04", "round": "港股 IPO（01879.HK）", "amount": "募资总额约 25.27 亿港元，净额约 23.77 亿港元", "lead": "基石投资者含阿里巴巴、GIC、贝莱德、富达国际、淡马锡、中移资本等 20 家", "src": "https://m.cnfin.com/yw-lb//zixun/20260428/4406000_1.html" } ], "public": true, "note": "2026-04-28 港交所上市，发行价 183.2 港元，首日收盘 886 港元；光互连收入约占八成，光计算收入占比小；官网域名凭记忆未核；成立时间各处有 2017/2018 两种口径，取 21 世纪经济报道 2017。" }, { "name": "BrainChip", "layer": "x2", "url": "https://brainchip.com", "hq": "拉古纳希尔斯（美国加州，ASX 上市）", "region": "海外", "desc": "全数字事件驱动神经形态芯片/IP Akida，边缘低功耗推理", "where": "相邻", "mode": "脉冲", "process_node": { "v": "22nm FD-SOI（GlobalFoundries，AKD1500）", "src": "https://brainchip.com/press/brainchip-announces-commercial-availability-and-production-shipments-of-akd1500-neuromorphic-processors/" }, "stage": { "v": "量产", "date": "2026-06", "note": "AKD1500 宣布商业可用并开始量产出货，同时进行工业与军规认证", "src": "https://brainchip.com/press/brainchip-announces-commercial-availability-and-production-shipments-of-akd1500-neuromorphic-processors/" }, "scene": [ "边缘", "端侧" ], "revenue": { "v": "US$1.89M，同比增长 374%", "date": "2025", "kind": "2025 年营收，年报，媒体转述", "src": "https://smallcaps.com.au/article/brainchip-licenses-akida-2-ip-to-edgeai-targeting-future-royalties" }, "total_raised": "累计融资未汇总（上市公司多次配售）", "funding": [ { "date": "2025-12", "round": "募资（机构配售）", "amount": "$25M（约 A$35M）", "lead": "未写明", "src": "https://www.design-reuse-embedded.com/news/202512068/brainchip-announces-25-million-usd-funding-ahead-of-ces-to-power-next-gen-edge-ai/" } ], "key_data": [ { "k": "AKD1500 算力与功耗", "v": "800 GOPS，低于 300 mW", "date": "2025-11", "src": "https://www.nasdaq.com/press-release/brainchip-unveils-breakthrough-akd1500-edge-ai-co-processor-embedded-world-north" } ], "customers": [ "Parsons（Blue Ridge Envisioneering，首批 1 万颗订单）", "EDGEAI（Akida 2 IP 授权）", "GlobalFoundries（制造伙伴）" ], "public": true, "note": "ASX: BRN，OTCQX: BRCHF；成立年份说法不一（早期为澳大利亚公司），没查到一手出处故省略；Akida 并非存内，是事件驱动数字架构，作相邻代表；2026-09 财报电话会提到 AKD1500 良率低于预期，仅见聚合页转述，未采。" }, { "name": "GrAI Matter Labs", "layer": "x2", "url": "https://www.eetimes.com/has-grai-matter-labs-been-snapped-up-by-snap-inc/", "hq": "巴黎", "region": "海外", "founded": 2016, "desc": "近传感的数字神经形态 SoC GrAI VIP，稀疏数据流加近存计算", "where": "相邻", "mode": "脉冲", "scene": [ "边缘" ], "total_raised": "累计融资约 $30M（EE Times 2024-03）", "key_data": [ { "k": "GrAI VIP 实测（公司曾向 EE Times 提供）", "v": "MobileNetV1-SSD 30fps 功耗 184 mW", "date": "2024-03", "src": "https://www.eetimes.com/has-grai-matter-labs-been-snapped-up-by-snap-inc/" } ], "owner": "2023-10 被 Snap（Snapchat 母公司）收购，金额未披露；EE Times 称据法媒报道，Snap 与 GrAI 均未回应置评", "note": "收购由 Bits&Chips 2024-02 首先报道（https://bits-chips.nl/article/grai-matter-labs-quietly-snapped-up-by-snap/ ）；官网域名没查到，url 用 EE Times 报道页。" }, { "name": "Innatera", "layer": "x2", "url": "https://www.innatera.com/about-us/", "hq": "代尔夫特", "region": "海外", "founded": 2018, "desc": "模拟/数字脉冲神经网络加速器，Pulsar 神经形态微控制器，常开传感", "where": "相邻", "mode": "脉冲", "process_node": { "v": "28nm（TSMC）", "src": "http://www.chinaaet.com/article/3000171782" }, "stage": { "v": "量产", "date": "2025-06", "note": "推出 Pulsar 量产版；此前 2024 年 T1 为量产前样品", "src": "https://www.eetimes.com/innatera-adds-more-accelerators-to-spiking-microcontroller/" }, "scene": [ "端侧" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2024-03", "round": "A 轮", "amount": "€15M，2024-06 扩至 $19.8M", "lead": "投资方为 Invest-NL Deep Tech Fund、EIC Fund、MIG Capital、Matterwave Ventures、Delft Enterprises，领投方未写明", "src": "https://www.eenewseurope.com/en/innatera-raises-e15m-for-neuromorphic-edge-ai/" } ], "key_data": [ { "k": "功耗与延迟（公司宣称，相对主流数字硬件）", "v": "延迟最多降 100 倍，能耗最多降 500 倍；常开场景低于 1 mW", "date": "2025-06", "src": "https://www.eetimes.com/innatera-adds-more-accelerators-to-spiking-microcontroller/" } ], "customers": [ "Aaroh Labs", "Cyran AI Solutions" ], "note": "扩轮 $19.8M 见公司 About 页；节点出自中文转述（芯智讯），未找到英文一手页；估值没查到。" }, { "name": "Intel Loihi / Hala Point", "layer": "x2", "url": "https://www.intel.com.br/content/www/br/pt/newsroom/news/intel-builds-worlds-largest-neuromorphic-system.html", "hq": "圣克拉拉", "region": "海外", "desc": "Loihi 2 神经形态研究芯片及 1.15B 神经元的 Hala Point 系统", "where": "相邻", "mode": "脉冲", "process_node": { "v": "Intel 4", "src": "https://www.sdxcentral.com/news/intel-unveils-115bn-neuron-neuromorphic-system-hala-point/" }, "stage": { "v": "研究", "date": "2024-04", "note": "Hala Point 部署在 Sandia 国家实验室，Intel 称其为研究原型", "src": "https://www.intel.com.br/content/www/br/pt/newsroom/news/intel-builds-worlds-largest-neuromorphic-system.html" }, "key_data": [ { "k": "系统规模", "v": "1,152 颗 Loihi 2，11.5 亿神经元，1,280 亿突触", "date": "2024-04", "src": "https://www.sandia.gov/news/publications/hpc-annual-reports/article/advancing-neuromorphic-computing-at-the-neural-exploration-and-research-lab/" }, { "k": "能效", "v": "运行传统深度网络超过 15 TOPS/W（INT8），最大功耗 2,600W", "date": "2024-04", "src": "https://www.sdxcentral.com/news/intel-unveils-115bn-neuron-neuromorphic-system-hala-point/" } ], "customers": [ "Sandia National Laboratories" ], "product_line_of": "Intel", "note": "founded 按 Loihi 首次公开年份应为更早，本轮没查到一手出处故省略（Loihi 2 为 2021 年发布）；领英帖子称 2026-01 有 Loihi 3（4nm、800 万神经元），无一手出处，未采。" }, { "name": "Sparxity（原 SpiNNcloud）", "layer": "x2", "url": "https://www.sparxity.ai/", "hq": "德累斯顿", "region": "海外", "founded": 2021, "desc": "基于 SpiNNaker2 众核芯片的类脑超算平台，TU Dresden 分拆", "where": "相邻", "mode": "脉冲", "scene": [ "数据中心" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2025", "round": "EIC Accelerator 混合融资", "amount": "€10M（欧盟）", "lead": "欧洲创新理事会（EIC）", "src": "https://eic.ec.europa.eu/impact/spinncloud_en" } ], "customers": [ "Sandia National Laboratories（部署 SpiNNaker 2）" ], "note": "额外补充的第 7 家类脑代表；2024 年向试点客户交付首批硬件、拿到首笔 VC（EIC 页）；Sandia 部署见 https://www.sandia.gov/news/publications/hpc-annual-reports/article/advancing-neuromorphic-computing-at-the-neural-exploration-and-research-lab/ ；官网域名没查到，url 用 EIC 页；工艺节点没查到。" }, { "name": "时识科技 SynSense", "layer": "x2", "url": "https://www.synsense.ai", "hq": "宁波", "region": "中国", "founded": 2017, "founded_month": "2025-01", "desc": "类脑感算一体：动态视觉 SoC Speck、Xylo 处理器、脑机接口信号芯片", "where": "相邻", "mode": "脉冲", "stage": { "v": "量产", "date": "2023", "note": "Speck 感算一体视觉 SoC 2023 年量产（百度百科口径）；2026-09 发布脑机接口信号采集芯片 Rigi 系列", "src": "https://baike.baidu.com/item/%E5%AE%81%E6%B3%A2%E6%97%B6%E8%AF%86%E7%A7%91%E6%8A%80%E6%9C%89%E9%99%90%E5%85%AC%E5%8F%B8/64919518" }, "scene": [ "端侧", "边缘" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2025-01", "round": "B+ 轮", "amount": "数千万元", "lead": "未写明", "src": "https://baike.baidu.com/item/%E5%AE%81%E6%B3%A2%E6%97%B6%E8%AF%86%E7%A7%91%E6%8A%80%E6%9C%89%E9%99%90%E5%85%AC%E5%8F%B8/64919518" }, { "date": "2026-06", "round": "B 轮", "amount": "数亿元", "lead": "国调二期协同发展基金、国海投资联合领投（恒泰华盛、华仓资本、南岭基金、和顺盈、宏力达跟投）", "src": "https://caifuhao.eastmoney.com/news/20260701204447075420410" } ], "key_data": [ { "k": "Speck 规格", "v": "约 32 万神经元，典型功耗低于 5 mW，分辨率 128×128", "date": "2023-06", "src": "https://www.synsense.ai/wp-content/uploads/2023/06/Speck-application-brochures-EN.pdf" } ], "note": "2017 年创立于苏黎世，宁波时识 2020-03 注册、总部迁中国（百度百科）；工艺节点没查到；2024 年收购 iniVation 一事只见公众号转述，未采；成立与迁址口径以百度百科为准，建议复核。" }, { "name": "灵汐科技 Lynxi", "layer": "x2", "url": "https://www.lynxi.com", "hq": "北京", "region": "中国", "founded": 2018, "desc": "清华类脑计算中心成果转化，众核类脑芯片领启 KA200 及板卡/服务器", "where": "相邻", "mode": "脉冲", "scene": [ "边缘", "数据中心" ], "total_raised": "累计融资未披露", "funding": [ { "date": "2024", "round": "B+ 轮", "amount": "未披露", "lead": "国鼎资本", "src": "https://api3.cls.cn/share/article/1665346?os=ios&sv=120&app=stib" } ], "note": "公司称 KA200 为存算一体、众核并行架构；工艺节点、量产状态、营收没查到一手；B 轮投资方含国家军民融合产业投资基金、中电海康（百度百科口径，未单独核）；2019 年成果登 Nature 封面。" } ] }