Skip to main content

32 posts tagged with "AI Chip"

AI chip industry dynamics and trends

View all tags

2026 训练 GPU 选型指南:从 H100 到 Rubin,预算百万到亿级怎么选

· 6 min read
Industry Research Team

训练选型的本质不是"买最强的",而是在预算、负载规模、软件生态和供货确定性之间找到平衡点。本文数据均取自本站芯片卡与定价库,帮你按预算分档做出决策。

一、先分级:你的训练负载属于哪一档​

训练负载大致分为三档,档位不同,选型逻辑完全不同:

  • 微调 / LoRA:单卡或 8 卡即可完成,7B–70B 模型为主。瓶颈在显存容量与单卡带宽,互联要求低,PCIe 机型甚至二手卡都能胜任。
  • 继续预训练 / 领域适配:数百卡规模,需要稳定的多卡互联与成熟的软件栈,容错与断点续训能力开始重要。
  • 从零预训练:千卡乃至万卡集群,互联带宽、机架方案(NVL72、Helios、CloudMatrix)和集群软件效率决定成败,预算通常以千万美元计。

二、按预算分档推荐​

预算档位推荐方案关键规格参考单价(站内定价库)
入门微调(数十万元级)H100 SXM / H200BF16 1,979 TFLOPS(稀疏)、80GB / 141GB HBM3e、NVLink 900 GB/s、700W 风冷H100 市场价约 $28,000;H200 约 $38,000
中等规模(数百万元级)B200 / AMD MI355XB200:192GB HBM3e、FP8 10 PFLOPS(稀疏)、NVLink 1.8 TB/s;MI355X:288GB HBM3E、FP8 dense 5 PFLOPS、UALink 600 GB/sB200 约 $42,000;MI355X 官方指导价 $22,000
旗舰预训练(千万到亿级)B300 Ultra / Rubin R200 / MI455XB300:288GB HBM3e、FP8 dense 7 PFLOPS;Rubin:288GB HBM4、22 TB/s、NVLink 6 3.6 TB/s;MI455X:432GB HBM4、23.3 TB/s、MXFP4 40.3 PFLOPSB300 约 $52,000;Rubin NVL72 机柜 350–400 万美元;MI455X 定价未公开(Helios 机柜约 500–550 万美元)
国产合规档昇腾 910C / 昇腾 950DT / 寒武纪思元690910C:BF16 800 TFLOPS、128GB HBM2e;950DT:FP8 约 1 PFLOPS、144GB 自研 HiZQ 4 TB/s;思元690:FP16 700+ TFLOPS、196GB HBM3910C 渠道价超 ¥11 万;950DT 超 ¥25 万;思元690 约 ¥17.5 万

几个要点值得展开:

  • 入门档首选 H100/H200 并不落伍。H100 的 BF16 稀疏算力 1,979 TFLOPS、700W TDP 风冷友好,生态最成熟;H200 显存加到 141GB、带宽 4.8 TB/s,微调 70B 级模型更从容。
  • 中等规模档 B200 与 MI355X 各有侧重。MI355X 用 288GB HBM3E 和 3nm 制程在显存容量上压过 B200,官方口径 FP8 dense 5 PFLOPS 也略胜,价格还便宜约四成;但 B200 的 NVLink 1.8 TB/s 是 UALink 600 GB/s 的三倍,千卡扩展时差距会显现。
  • 旗舰档看机架而不是单卡。B300 Ultra 必须液冷(TDP 1,400W),Rubin R200 的 Vera Rubin NVL72 机柜 FP4 算力约 3.6 EFLOPS、总显存 20.7TB;MI455X 的 Helios 机架 72 卡 31TB HBM4、机柜功耗 225–245kW。这一档的采购单位是"机柜",还要配套液冷与供电改造。

三、关键参数怎么读​

  • FP16/BF16 算力:注意 dense 与 sparse 口径差一倍。H100 的 1,979 TFLOPS 是稀疏值,dense 约 989 TFLOPS;比较时务必统一口径。
  • 显存容量:从 H100 的 80GB 到 B300 的 288GB 再到 MI455X 的 432GB,容量决定了能训多大模型、能开多大 batch。训练 70B 模型 FP16 权重本身就要约 140GB,加上优化器状态后通常需要多卡切分。
  • 互联带宽:NVLink 4.0(900 GB/s)→ NVLink 5(1.8 TB/s)→ NVLink 6(3.6 TB/s);AMD 走开放路线的 UALink 600 GB/s;昇腾为 HCCS 784 GB/s 与灵衢 2 TB/s。张量并行通信密集,互联带宽直接决定扩展效率。
  • 机架方案:GB200/GB300 NVL72、Helios、CloudMatrix 384 是三家给出的系统级答案。单卡强不等于集群强,采购前要确认整机柜方案、网络与液冷的完整交付能力。

四、软件生态权重:CUDA vs ROCm vs CANN​

  • CUDA:十八年积累,算子覆盖最全,从 H100 到 Rubin 全线兼容,迁移成本最低。
  • ROCm:AMD 官方口径承认成熟度仍落后 CUDA 约 2–3 年,但开放生态对云厂商和主权 AI 项目有吸引力,MI355X/MI455X 的客户名单里已有 Microsoft Azure、Meta、Oracle 等。
  • CANN:昇腾 950 系列配套 CANN Next 提供 CUDA 兼容层,约 80% 的标准 PyTorch 代码仅需配置修改即可迁移;2026 年 10 月 DeepSeek 与华为还联合发布了面向昇腾 950 的开源编程工具与库。

经验法则是:微调档闭眼选 CUDA;中等规模档如果显存需求突出、团队有 ROCm 经验,MI355X 值得评估;合规档则要在选型时就把迁移人力写进预算。

五、供货与价格现实​

2026 年的选型绕不开 HBM:

  • 涨价传导:HBM 成本攀升直接推高国产卡报价——昇腾 950DT 渠道指示价已超 ¥25 万(较两个月前上涨 20%–50%),昇腾 950PR 突破 ¥8 万,寒武纪等同期上调 20%–30%。
  • 云租是一条现实路径:H100 云租约 $2.00/GPU·小时,B200 约 $5.87/GPU·小时,B300 的 AWS 8 卡实例约 $11.70/GPU·小时(Nebius 口径约 $9.50/GPU·小时)。训练周期短的项目,租往往比买划算。
  • 旗舰卡要排队:Rubin R200 已于 2026 年 6 月全面量产,但自建机房用户要到 2027 年一季度才能采购;MI455X 也需到 2026 年下半年至 2027 年上半年才规模爬坡。急用算力,H100/H200/B200 的现货市场仍是主力。

想逐项对比更多型号(包括 Google TPU、Intel Gaudi 4、百度昆仑芯等),请查看站内完整对比表;每张卡的详细规格可进入对应芯片卡页面查看。

小结​

2026 年的训练 GPU 市场第一次出现了"四条路线":CUDA 旗舰(Rubin/B300)、开放生态(MI355X/MI455X)、推理专用硬件和国产合规方案。选型的顺序应该是:先定负载档位,再按预算筛卡,然后核算互联与机架方案,最后把软件迁移和供货周期计入总成本。单卡峰值参数只讲了一半的故事,剩下的另一半在集群和生态里。

台积电海外建厂与 AI 芯片供给:亚利桑那、熊本与德累斯顿

· 5 min read
Industry Research Team

每一颗AI加速器都绕不开台积电的晶圆。当AI需求以指数级膨胀,"把所有先进制程放在一座岛上"的地缘脆弱性就成了全球供应链最敏感的神经。台积电的答案是把厂搬出去——但搬出去的产能,和留在岛内的产能,并不是同一种产能。

扩建速度翻倍:18座工厂的全球棋局​

在2026年技术论坛上,台积电运营与先进技术工程副总经理田博仁表示,面对AI与高性能运算的强劲需求,公司正以过去两倍的速度加速扩建,全球范围内将新建与改造共计18座半导体工厂,其中包含5座先进封装厂。2025至2026年间,台积电每年可新建约9座工厂,较2017至2024年每年约4座的节奏翻倍。台积电同时预计,2022年至2026年间AI加速器晶圆需求将增长11倍,2纳米及更先进制程产能在2026至2028年的年复合增长率达70%。

海外基地制程定位最新进展
亚利桑那一厂4纳米2024年量产,2026年产能提升1.8倍
亚利桑那二厂3纳米主体完工,2027年下半年投产
亚利桑那三厂2纳米级2026年5月封顶
熊本一厂22与28纳米已量产,良率追平台湾厂区
熊本二厂升级至3纳米2025年动工建设
德累斯顿ESMC28与22、16与12纳米主体完工,目标2027年底投产

亚利桑那:AI 芯片的美国制造心脏​

亚利桑那是这场海外扩张的绝对重心。2026年5月,台积电核准在不超过200亿美元的额度内增资美国子公司,作为去年公布的1650亿美元全球扩建计划的提速动作。英伟达、AMD、苹果等主要客户已预订其产能,台积电还计划在现有基础上继续扩容,考虑使亚利桑那的晶圆厂总数达到8座、封装厂达4座,形成"超级晶圆厂"集群。

关键的战略指标是:未来台积电在美国的产能将占其全球2纳米及更先进制程的约30%。这对AI芯片供给意味着什么?英伟达与AMD最尖端的加速器晶圆将有三分之一产自美国本土,出口管制、关税与地缘风险对这部分产能的扰动模式将完全不同。值得注意的是,亚利桑那厂区已从此前累计约12.5亿美元的爬坡亏损(2021至2024年,据公开报道)转为单季盈利——2026年第一季度盈利约188亿元新台币,超过2025年全年总和。

熊本与德累斯顿:两种完全不同的海外逻辑​

日本熊本与德国德累斯顿的角色定位截然不同。熊本一厂目前量产22与28纳米,台积电披露2026年该厂28与22纳米产出将达2025年的20倍;二厂原本规划6纳米与7纳米,因AI需求强劲已升级为3纳米制程——这是熊本首次触碰先进制程。不过2026年7月底熊本发生的强震一度迫使JASM停工,虽然厂体无损,但也直观展示了"多元化"本身也会遭遇多元化风险。

德累斯顿则压根不是为AI而生。由台积电与博世、英飞凌、恩智浦合资的ESMC聚焦28与22纳米及16与12纳米制程,面向汽车与工业客户,欧洲芯片法案承担了接近一半的资本开支。它解决的是欧洲"汽车芯片完全依赖亚洲"的焦虑,而非AI算力的供给问题。

研判:先进制程的"去中心化"才刚起步​

读懂台积电海外布局的关键,是认清一个事实:目前最尖端的AI芯片制造主力依然留在台湾——2纳米大规模量产由新竹与高雄厂承担,海外产能要到2027年后才逐步接棒先进节点。海外建厂买的是地缘保险与客户信任,付出的则是建厂成本数倍于台湾、工程师与文化磨合的代价。对AI产业而言,未来三到五年会形成一个分层供给结构:美国负责最尖端制程的三成产能,日本承接先进与成熟制程的过渡带,欧洲专注车规与工业。供应链韧性在提升,但供给的定价权,依然牢牢握在新竹的排产会议桌上。

晶圆级 vs 芯粒:两条极限算力路线的碰撞

· 4 min read
Industry Research Team

当整个行业都在拥抱芯粒时,Cerebras 却坚持把一整片 12 英寸晶圆做成一颗芯片。2026 年,这条"离经叛道"的路线迎来了 CS-4 量产与纳斯达克上市的双重节点。晶圆级与芯粒,究竟谁代表极限算力的未来?

晶圆级:把晶圆当芯片用​

Cerebras 的 WSE-3 是迄今最大的商用芯片之一:整片 300mm 晶圆不切割,4 万亿晶体管、约 90 万个 AI 核心通过片上网络互联,片上 SRAM 达 44GB,片内带宽据厂商数据高达 21 PB/s——这是"消灭数据搬运"的极致设计,芯片面积达到约 4.6 万平方毫米。

2026 年 8 月,Cerebras 发布了基于三片 WSE-3T 晶圆的 CS-4 系统:AI 算力 750 PFLOPS,晶圆间带宽 129.6 PB/s,宣称在 GPT-OSS-120B 上实现单用户每秒超过 4400 个 tokens 的生成速度。Hot Chips 2026 上,公司进一步公布了 Nexus 背包式供电架构——直接从铜排背面给晶圆供电,省去 PCB 与基板的传输损耗,使同等硅片性能翻倍——以及 2027 年 WSE 新一代硅片与集成堆叠 DRAM 的 CS-6 规划。资本市场也给出认可:据公开报道,Cerebras 以每股 185 美元登陆纳斯达克,募资约 55 亿美元,并与 OpenAI 签订了至 2028 年最高 750 兆瓦的算力协议。

芯粒:良率与成本的最优解​

与晶圆级的"整片浇筑"相对,芯粒路线选择"化整为零":把计算、缓存、I/O 拆成小芯粒分别流片,再用先进封装拼合。英伟达 Rubin R200 的 2 颗计算 die 加 2 颗 I/O die 加 8 堆栈 HBM4 组合、AMD MI400 系列、以及华为昇腾等国产主力芯片,走的都是这条路。芯粒的优势在于小 die 良率高、可混搭制程、封装组合灵活,是当前商业化的绝对主流。

维度晶圆级(Cerebras WSE)芯粒(Rubin、MI400 等)
面积哲学整片晶圆即芯片,约 4.6 万平方毫米多颗小 die 拼合封装
片上存储44GB SRAM,片内带宽 PB/s 级依赖 HBM,单卡容量 288GB 级
强项场景超低延迟推理、Agent 实时交互训练、大显存推理、通用部署
软件生态独立 SDK,需专门适配CUDA、ROCm 生态广泛覆盖
商业模式云服务与超大客户直销通用硬件大规模出货

两条路线的适用边界​

晶圆级的瓶颈同样明显:44GB SRAM 装不下千亿参数模型,需要外部流式供数或多晶圆互联扩展;单系统功耗高达 25 千瓦级,必须定制工业液冷;客户集中度极高,据招股材料,其阿联酋大客户曾占公司收入的 85% 以上。它注定是一条"特种尖刀"路线,专注低延迟推理、超长上下文与科研超算等小众高价值场景——毕竟 Agent 应用是"没有耐心的",每秒数千 tokens 的单用户体验正是 GPU 集群最难优雅解决的痛点。

芯粒则是规模化基本盘:性能上限虽受封装互联约束,但良率、成本与生态兼容性全面占优,能覆盖从单卡到超节点的全部市场。这也是国产算力选择以芯粒为主攻方向的根本原因——成熟制程产能加自主先进封装,是当前产业禀赋下性价比最高的突围路径。

结语:碰撞之外,各守一极​

晶圆级与芯粒并非零和竞争,而是从两端同时凿穿"访存墙"的两种尝试:前者把存储距离压到零,后者用先进封装逼近光罩物理极限。2027 年之后,随着延迟敏感型推理负载占比上升,晶圆级路线的商业空间可能继续扩大;而在通用算力市场,芯粒的地位短期内无可撼动。真正的输家,也许是停留在单片巨 die 时代、两头都不沾的中间路线。

xAI 与特斯拉 Dojo 2:非典型玩家的自研芯片路线

· 4 min read
Industry Research Team

在超大规模云厂商与芯片大厂之外,马斯克的两家公司给出了自研芯片的另一种打开方式:一家砍掉自研训练超算、押注推理芯片量产,一家干脆不造芯片、把租来的 GPU 堆成全球最大集群。

Dojo 的终结与 AI5 的转向​

2025 年 8 月,马斯克在社交平台确认解散特斯拉 Dojo 训练超算团队,并给出一句著名判断:Dojo 2 是一条"演化死胡同"。他解释称,当所有路径都汇聚到 AI6 芯片时,特斯拉没有理由同时维持两套差异巨大的芯片设计。这座投入了 5 亿美元的布法罗 Dojo 设施随之失去原有使命。

转向的落点是车载推理芯片。据公开信息,AI5 芯片于 2026 年 4 月完成流片定稿,由台积电亚利桑那厂与三星得州厂双源代工,2027 年年中启动大规模量产,优先服务 Cybercab 与 Optimus 项目;AI6 处于早期开发阶段,目标性能较 AI5 翻倍,定位同时覆盖车载推理、机器人控制与大规模 AI 训练。2026 年 5 月,特斯拉又以 Dojo 2 之名发布了一款面向 Cybercab 的自动驾驶芯片,据业界报道宣称可在 45W 功耗下提供 900 TFLOPS 算力,8 月起批量装车。

阶段关键动作时间线
Dojo 一代D1 芯片加英伟达 GPU 混合集群2023 年上线
项目关停团队解散,D2 芯片搁置2025 年 8 月
AI5台积电与三星双源代工2026 年流片,2027 年中量产
Dojo 2 命名车载推理芯片,45W 下 900 TFLOPS2026 年 5 月发布
AI6目标性能翻倍计划 2028 年量产

"Dojo 3" 的新含义:用通用芯片攒超算​

马斯克为 Dojo 这个名字保留了另一种存续方式:既然 AI5 与 AI6 在推理之外"至少还不错"地能做训练,那么把大量 AI6 芯片集成到单块电路板上,就能构成训练集群,还能把网络布线的复杂度与成本降低几个数量级。他甚至直接表示,"或许可以把它称为 Dojo 3"。这实质上宣告了专用训练芯片路线的终结——用一条高量产的车规芯片产线,摊薄训练集群的芯片成本。

与特斯拉形成有趣对照的是 xAI。这家公司几乎完全放弃了芯片自研,而是以惊人的速度从英伟达采购和租借 GPU,在孟菲斯建成了全球首个吉瓦级训练集群 Colossus 2,并逐步把富余算力对外出租,谷歌、Anthropic 均已成为其大客户。马斯克体系内由此形成分工:xAI 造算力工厂,特斯拉造端侧芯片,训练与推理各取所需。

非典型路线的启示​

特斯拉与 xAI 的组合揭示了自研芯片的另一种经济学:当端侧场景(车辆、机器人)足够确定且出货量足够大时,自研芯片的核心价值在于规模与迭代速度——马斯克甚至希望把未来芯片的设计周期压缩到 9 个月;而当需求爆发快于自研节奏时,直接采购现成 GPU 抢时间,比造芯片更理性。三星与特斯拉之间 165 亿美元的代工协议,也说明车载 AI 芯片已成为先进制程的重要买家。

小结​

Dojo 的落幕并非自研芯片的失败,而是一次清醒的止损:把有限资源从长周期的训练超算转向出货量以百万计的端侧推理芯片。在"要不要自研"这个问题上,马斯克给出了一个与云厂商截然不同却同样自洽的答案——自研与否,取决于你的芯片要装在哪里、装多少颗。

燧原科技登陆科创板:开盘暴涨188%、市值1764亿元,国产AI芯片资本化再下一城

· 5 min read
Industry Research Team

开盘暴涨188.37%、市值1764.44亿元——燧原科技的科创板首秀,为国产AI芯片的资本化浪潮再添一把火。

首日表现:高开高走​

2026年9月11日,上海燧原科技股份有限公司正式在上海证券交易所科创板挂牌上市,股票代码688801.SH。开盘价410元/股,较发行价暴涨188.37%,市值达到1764.44亿元;盘中股价最高触及468元/股,涨幅超过229%。

对于一家AI芯片公司而言,这样的首日表现并不常见。它反映的是二级市场对国产AI算力标的的稀缺性定价:在AI算力需求爆发、国产替代加速的双重叙事下,能够提供训练+推理完整产品线的芯片公司,几乎是科创板最炙手可热的资产类别。

指标数值
上市日期2026年9月11日
股票代码688801.SH
上市板块上交所科创板
开盘价410元/股
开盘涨幅188.37%
盘中最高价468元/股(涨幅超229%)
上市初期市值1764.44亿元
重要股东腾讯(浮盈超300亿元)

腾讯作为燧原科技的重要股东,在这场上市中浮盈超过300亿元。从早期一路陪跑到科创板敲钟,腾讯对燧原的持续投入,是互联网大厂深度绑定国产AI芯片公司的典型案例,也印证了一级市场对国产算力赛道的长期押注。

产品底色:云燧训练/推理卡与集群方案​

资本故事之外,燧原的产品线是其估值的底色。燧原科技的核心产品围绕"邃思"DTU芯片展开,基于该芯片打造云燧系列训练加速卡与推理加速卡,并延伸出数据中心级集群方案,覆盖从单卡到大规模部署的完整需求链条。

以燧原T21为例,该卡搭载64GB HBM2E显存,面向大模型时代的显存密集型负载。在HBM容量与带宽成为AI加速卡竞争焦点的当下,大容量HBM配置是衡量产品竞争力的硬指标之一。而集群方案的完整度——从单卡、整机到组网调度——则决定了芯片公司能否真正承接云厂商与行业客户的数据中心级订单。

国产AI芯片的集体资本化​

燧原上市并非孤立事件,而是国产AI芯片企业批量登陆资本市场的最新一站。在此之前,板块格局已经初步成形:

公司上市地/代码
寒武纪科创板
摩尔线程科创板
沐曦科创板(688802.SH)
天数智芯港股(09903.HK)
燧原科技科创板(688801.SH,2026年9月11日)

从训练芯片到推理芯片,从GPU路线到GPGPU、DTU等多技术路线,国产AI芯片公司正在以"科创板为主阵地、港股为补充"的格局完成集体资本化。燧原开盘188%的涨幅,某种程度上是市场对整个板块稀缺性与成长性的集中定价——每一个新上市标的,都在为后来者校准估值坐标。

"融资-流片-上市-再投入"的飞轮​

AI芯片是典型的重资产、长周期行业:一次先进制程流片动辄数亿美元级投入(行业内公认的成本结构),大规模量产还需要持续的产能保障与人员扩张。没有持续的资本供给,技术追赶很难维持节奏。

此番IPO募资,燧原将投向研发与产能两个方向:前者保证下一代芯片的技术迭代与追平能力,后者保证规模交付能力。由此,国产AI芯片行业正在形成一个清晰的飞轮:

  1. 融资:一级市场资金支撑芯片流片与团队建设;
  2. 流片与量产:产品落地,进入云厂商与行业客户供应链;
  3. 上市:借助资本市场完成价值发现与再融资通道建设;
  4. 再投入:募资投向下一代研发与产能,进一步缩小与海外厂商的差距。

寒武纪、摩尔线程、沐曦、天数智芯与燧原的接连上市,让这个飞轮转得越来越快。资本市场的持续输血,正在成为国产AI芯片追赶进程中不可或缺的助推器。

小结​

燧原科技科创板首日开盘暴涨188.37%、市值1764.44亿元、盘中最高涨幅超229%,腾讯浮盈超300亿元——这既是资本市场对国产AI芯片稀缺标的的热烈回应,也是"融资-流片-上市-再投入"产业飞轮加速运转的缩影。在HBM紧缺与AI算力需求爆发的双重背景下,国产AI芯片的资本化窗口仍在敞开。后续燧原能否将高估值转化为持续的营收与利润,将是检验这场资本盛宴成色的关键。

HBM4 从送样竞速进入量产爬坡竞速:SK 海力士率先交付 Rubin、美光产能翻倍、三星直逼王座

· 6 min read
Industry Research Team

9 月第一周,存储行业传来三组关键信号:SK 海力士面向英伟达 Vera Rubin 平台量产交付全球首款 12 层 HBM4;三星 Q2 HBM 份额飙升至 33%、直逼王座;美光宣布年底 HBM 月产能翻倍至约 10 万片晶圆。 HBM4 的竞争,已经从"谁能造出来"变成"谁能最快量产爬坡、谁绑定的客户最深"。


1. 三巨头战报:领跑者、追赶者与搅局者​

维度SK 海力士三星电子美光
HBM4 关键节点2026 年中向 Vera Rubin 量产交付 12 层 HBM4(全球首款完整质量认证)2026 年 2 月全球首家大规模量产出货2026 Q2 量产 12 层 HBM4,批量出货
DRAM 核心裸片1b(第五代 10nm 级)1c(第六代)—
Base Die 代工台积电 12nm自家 4nm拟转台积电(1γ 工艺代)
2026 HBM4 份额预测~54%~28%~18%
Q2 HBM 营收份额50%(环比 -14pct)33%(环比 +12pct)18%
HBM 产能~15–20 万片/月~15–20 万片/月年底冲刺 10 万片/月(现为 4–5 万)

三个关键读数:

  • 三星的反扑是真的:从去年 Q2 份额低至 15%(落后美光 6 个百分点)到今年 Q2 的 33%,三星用一年时间把与 SK 海力士的差距缩小到 17 个百分点。三星预计 HBM4 将占其下半年 HBM 收入的 60% 以上,且现有 HBM4 产能已被客户预订售罄,2026 年 HBM 总产能计划提升约 50%;
  • 美光的追赶也是真的:CEO 梅赫罗特拉透露 12 层 HBM4 爬坡速度约为 HBM3E 的两倍、累计收入已超 10 亿美元;年底若如期达成 10 万片月产能,与两强的差距将缩小一半;
  • SK 海力士的护城河依然深:与英伟达签署多年期 HBM/DRAM 供应协议、率先完成 2027 年 HBM4 价格谈判,在英伟达 HBM4 采购中占比约 60%(乐观情形 70%)。

2. 变局核心:封装比 DRAM 制程更值钱​

HBM4 与 HBM3E 时代最大的不同,是胜负手从存储颗粒转向了封装与 Base Die:

  • I/O 通道数从 1024 条翻倍至 2048 条,单颗 12 层封装吞吐超 2TB/s,能效较前代提升逾 40%;
  • Base Die 首次引入晶圆代工厂先进逻辑工艺——SK 海力士用台积电 12nm,三星用自家 4nm,美光计划在 1γ(首次导入 EUV)世代转由台积电代工;
  • 存储原厂与代工厂的协作深度,第一次成为决定出货速度的关键变量——这正是 Hot Chips 2026 上"三星 HBM Base Die 逻辑工艺化"议题的产业注脚。

SEMI 数据显示,2026 年全球 HBM 市场规模预计增长 58% 至 546 亿美元,约占 DRAM 市场四成。行业预计 HBM4 销售占比将在 2026 年 Q4 正式超越 HBM3E,成为市场主流。

3. 客户端信号:Rubin 加码,Rubin CPX 重生​

需求端同样在 9 月出现重要变化:

  • Vera Rubin 全面量产(详见本站专文):每颗 Rubin GPU 搭载 288GB HBM4、带宽 22TB/s,NVL72 整柜 75TB 快速内存——HBM4 供给直接决定 Rubin 出货上限,这也是英伟达把"先进晶圆与 HBM 供应"列为当前第一约束的原因;
  • Rubin CPX 项目重启:据 CFM 闪存市场 9 月 2 日简讯,英伟达重启此前搁置的 Rubin CPX 机柜项目,内存规格由 128GB GDDR7 改为 168GB HBM4,机架改为独立 MGX ETL 设计,客户可选 64/128/192/256 颗 GPU 配置——推理专用卡也全面转向 HBM4,进一步放大了 HBM4 需求;
  • 供给协议长期化:英伟达已与 SK 海力士、美光签署多年期 HBM 及 DRAM 供应协议,锁供给、锁价格、锁产能成为巨头标配动作。

4. 下半场:HBM4E 已经鸣枪​

下一代产品的竞争在 2026 年同步开启:

  • SK 海力士:HBM4E 送样提前至 2026 年年中(原计划下半年),COMPUTEX 2026 已展出 12 层样品,单引脚速率最高 16Gbps、单堆栈带宽约 4TB/s,计划 2027 年量产;
  • 三星:2026 年 Q2 起向头部客户交付业界首批 HBM4E 样品(48GB),目标 2027 年拿下 HBM4E 市场 50% 以上份额;
  • 美光:HBM4E 预计 2027 年量产,首批样品采用 1γ 制程 DRAM。

5. 对采购方与投资观察者的启示​

  • 供应商组合成为第一优先级采购变量:同样买 Rubin 平台,采用哪家 HBM 供应商的整机,交付周期可能相差数月;
  • 国产链条的机会窗口:HBM 供不应求叠加出口管制,长鑫等国产存储玩家的 HBM 进展值得持续跟踪——国产 AI 芯片(昇腾 950 系列等)对 HBM 的需求同样在放量;
  • 跟踪三个领先指标:头部云厂商 Rubin 机架采购量、HBM 设备供应商订单(美光已加大 PO)、HBM4E 送样→量产的时间差。

相关链接​

参考资料​


本文基于 2026 年 9 月初 Counterpoint Research、CFM 闪存市场及韩媒报道整理。份额与产能数据均为研究机构/供应链预估口径,实际以各原厂财报为准。

Huawei Ascend 910C Deep Dive: Specs, Deployment, and Full Performance Overview

· 8 min read
Industry Research Team

Huawei Ascend 910C (Ascend 910C), Huawei's third-generation Ascend AI chip, adopts innovative dual-die (Chiplet) packaging and began mass supply in May 2025, becoming the backbone of domestic AI compute.

This article comprehensively analyzes this domestic flagship AI chip across four dimensions: technical specs, deployment cases, performance comparison, and market positioning.


1. Core Technical Specifications​

1.1 Chip architecture and process​

ItemParameter
ArchitectureDa Vinci (dual-die packaging)
ProcessSMIC N+2 (7nm-class)
PackagingChiplet (2× Ascend 910B compute dies)
Transistors~53 billion
Die size~800mm² (estimated)

Technology highlights:

  • Dual-die Chiplet packaging integrates two 910B chips, breaking the single-die yield bottleneck
  • Centerless I/O die design lets the two compute dies interconnect directly, reducing communication latency
  • SMIC N+2 process delivers 7nm-class performance with a controllable, autonomous supply chain

1.2 Compute performance​

PrecisionComputeReference
BF16800 TFLOPS~60% of NVIDIA H100
FP16~800 TFLOPSClose to H100 at same precision
INT8~1600 TOPSClear inference advantage
FP32Not disclosedTraining mainly uses BF16/FP16

Performance characteristics:

  • 800 TFLOPS at BF16, a new domestic AI chip compute benchmark
  • ~2× compute over 910B (dual-die stacking + architecture optimization)
  • No FP8 precision support (NVIDIA Blackwell's strength)

1.3 Memory and interconnect​

ItemParameter
HBM typeHBM2E (8 stacks)
Memory capacity~128 GB (combined dual-die)
Memory bandwidth784 GB/s
Interconnect protocolHuawei AscendLink (in-house)
Interconnect bandwidth400 GB/s unidirectional (800 GB/s bidirectional)

Memory advantages:

  • 128GB capacity supports full-pipeline training of hundred-billion-parameter models
  • 784 GB/s is a high-end configuration among HBM2E solutions
  • In-house AscendLink protocol supports 384-chip all-optical interconnect

1.4 Power and energy efficiency​

ItemParameter
TDP (dual-die)~310 W
Energy efficiency (BF16)~2.58 TFLOPS/W
vs. H100~45% of H100's power, comparable energy efficiency

Energy efficiency advantages:

  • At equal compute, significantly lower power than NVIDIA H100 (700W)
  • 7nm-class process, ~30% better energy efficiency than 910B
  • Suited to large-scale cluster deployment, reducing data center PUE pressure

2. Key Deployment Cases​

2.1 CloudMatrix 384 SuperNode​

System specs:

ItemConfiguration
Chip count384 Ascend 910C
Cabinets16 (12 compute + 4 network)
Total HBM~49 TB (128GB × 384)
InterconnectAll-optical mesh network
Optical modules6,912 LPO optical modules
System BF16 compute~300 PFLOPS

Performance comparison:

  • CloudMatrix 384's total BF16 compute exceeds NVIDIA GB200 NVL72 (72× B200)
  • In large-model training, 384-chip 910C linear scaling efficiency reaches 85%+
  • Supports smooth scaling to ten-thousand-card clusters for ultra-large training

Deployment progress:

  • As of June 2026, over 500 CloudMatrix 384 SuperNodes deployed
  • Key customers: China Telecom, China Mobile, China Unicom, Huawei Cloud, iFlytek
  • Scenarios: large-model training, smart customer service, autonomous-driving simulation, scientific computing

2.2 DeepSeek-V4-Pro full-parameter post-training​

Breakthrough significance:

On June 5, 2026, the AI training platform of Shenzhen Hetao College — together with Harbin Institute of Technology (Shenzhen), Shenzhen Big Data Research Institute, Huawei, and Shenzhen Zhicheng AI Compute Platform — completed full-parameter post-training of the 1.6-trillion-parameter DeepSeek-V4-Pro large model on an Ascend 910C compute cluster.

Technical highlights:

  • Among the world's first to run full-parameter post-training of a trillion-parameter model on a domestic compute platform
  • Validates Ascend 910C maturity in ultra-large-model training
  • Proves domestic AI chips now have the capability to replace imported chips

Performance data (official disclosure):

  • Training throughput: ~60% of an H100 cluster (BF16 precision)
  • Memory utilization: 92% (128GB HBM2E capacity advantage)
  • Interconnect efficiency: 384-chip linear scaling efficiency 85%+
  • Stability: 30 consecutive days of training with no failures

2.3 Commercial deployment cases​

Case 1: A provincial big-data center (300 P FLOPS compute center)​

  • Scale: 300 P FLOPS AI compute (~1,000× 910C)
  • Scenarios: government large model, city brain, smart transportation
  • Deployment: September 2025
  • Investment: ~¥200M (120 servers)

Case 2: Huawei Cloud AI training platform​

  • Chips: over 10,000 Ascend 910C
  • Customers served: over 500 enterprises
  • Model support: Pangu large model, third-party open-source models (LLaMA, ChatGLM, etc.)
  • Global deployment: China, Southeast Asia, Middle East, Latin America

Case 3: iFlytek smart education​

  • Scale: 256 Ascend 910C
  • Scenarios: smart-education large model, speech recognition, machine translation
  • Performance: 90% faster training than 910B

3. Performance Comparison Analysis​

3.1 vs. NVIDIA H100​

ItemAscend 910CNVIDIA H100Notes
BF16 compute800 TFLOPS~1,300 TFLOPS910C ~60% of H100
HBM capacity128 GB80 GB910C +60%
HBM bandwidth784 GB/s3.35 TB/sH100 clear bandwidth lead
TDP310 W700 W910C only 45% of H100 power
Process7nm (SMIC N+2)4nm (TSMC)H100 more advanced
Software ecosystemCANN (CUDA-compatible)CUDAH100 more mature
SupplyChina autonomousExport-controlled910C no supply-chain risk

Conclusion:

  • In raw compute, 910C is ~60% of H100
  • In memory capacity, 910C leads by 60%, suited to large-model training
  • In energy efficiency, 910C clearly outperforms H100
  • In supply chain security, 910C wins outright

3.2 vs. Ascend 910B​

ItemAscend 910CAscend 910BImprovement
ArchitectureDual-die ChipletSingle die—
BF16 compute800 TFLOPS~400 TFLOPS+100%
HBM capacity128 GB64 GB+100%
TDP310 W310 WFlat (single-die power)
ProcessSMIC N+2SMIC N+2Same
Yield~40%~30%+33%

Conclusion:

  • 910C's dual-die packaging doubles compute and memory capacity
  • Yield up from 910B's 30% to 40%, lowering manufacturing cost
  • At equal power, 100% performance gain, significantly better energy efficiency

3.3 Inference performance (DeepSeek measured)​

Test environment:

  • Model: DeepSeek-V3 (671B parameters)
  • Hardware: Ascend 910C vs NVIDIA H100
  • Precision: BF16
  • Batch size: 64

Results:

MetricAscend 910CNVIDIA H100Ratio
Inference speed (tokens/s)8,50014,20060%
First-token latency (ms)12085141%
Power (W)31070044%
Cost (¥10k/card)~10~1856%

Conclusion:

  • 910C inference speed is 60% of H100, but power only 44%
  • In cost-sensitive scenarios, 910C's cost-performance advantage is clear
  • For China-market localization needs, 910C is the only option

4. Market Positioning and Competitive Advantages​

4.1 Target markets​

Core markets:

  1. Chinese government and SOEs: localization, data security, autonomy
  2. Large-model startups: cost-sensitive, high compute demand
  3. Operators and cloud providers: large-scale deployment, high efficiency requirements
  4. Research and education: ultra-large-scale computing, talent development

Edge markets:

  1. Autonomous driving: end-to-end large-model training
  2. Smart healthcare: medical imaging, drug discovery
  3. Fintech: risk control, robo-advisory

4.2 Competitive advantages​

AdvantageDescription
AutonomySMIC N+2 process + Huawei in-house architecture, no supply-chain risk
Large memory128GB HBM2E, full-pipeline training of hundred-billion-parameter models
High energy efficiency310W TDP delivers 800 TFLOPS, close to H100 efficiency
System scalingCloudMatrix 384 SuperNode, total compute exceeds GB200 NVL72
Software ecosystemCANN CUDA-compatible, lower migration cost
Cost advantage~¥100k/card, ~44% cheaper than H100

4.3 Weaknesses and improvement directions​

WeaknessImprovement direction
Single-chip computeNext-gen 910D to adopt 3nm, target doubling
HBM bandwidth950 series to adopt in-house HBM (HiBL 1.0), bandwidth to 4 TB/s
Software ecosystemContinued CANN + MindSpore investment, expand developer community
ProcessDeep cooperation with SMIC to ramp N+3 (5nm-class)

5. 2026 Shipment Plan and Market Forecast​

5.1 Shipment plan​

PeriodShipmentsCumulativeKey customers
2025 Q2-Q4200k200kHuawei Cloud, China Telecom
2026 Q1-Q2300k500kChina Mobile, China Unicom, iFlytek
2026 Q3-Q4300k800kGovernment projects, large-model startups
20271,000k1,800kGlobal market (Southeast Asia, Middle East, Latin America)

Capacity bottleneck:

  • SMIC N+2 capacity ~100k wafers/month, Ascend 910C ~30% of that
  • 2026 plan of 800k chips needs ~400k wafers, requiring 80%+ utilization
  • Huawei prioritizes 910C capacity via deep SMIC cooperation

5.2 Market forecast​

China AI chip market (2026):

  • Total: ~¥50B
  • Domestic share: ~35% (¥17.5B)
  • Ascend 910C share: ~60% (¥10.5B, ~800k chips)

Global AI chip market (2026):

  • Total: ~$200B
  • Huawei share: ~5% ($10B)
  • Growth drivers: China-market localization + Belt and Road exports

6. Summary and Outlook​

6.1 Core conclusions​

  1. Ascend 910C is a milestone domestic AI chip, with comprehensive breakthroughs in compute, memory, energy efficiency, and system scaling
  2. CloudMatrix 384 SuperNode proves domestic chips can replace imported ones
  3. DeepSeek-V4-Pro training success validates 910C maturity in ultra-large-model training
  4. 800k chips shipped in 2026, projected 60% of China's AI chip market

6.2 Future outlook​

Short term (2026-2027):

  • 910C continues ramping, shipments exceed 1,000k
  • CloudMatrix 384 deployments over 1,000 units
  • Software ecosystem (CANN + MindSpore) maturity approaches 70% of CUDA

Medium term (2028-2029):

  • Next-gen 910D mass production, 3nm process, target 1.6 PFLOPS BF16
  • 950 series (PR/DT) becomes inference-market mainstay, share over 30%
  • 960/970 launch, N+3 process, supports trillion-parameter models

Long term (2030+):

  • Huawei Ascend series becomes TOP 3 of the global AI chip market
  • Domestic AI chips exceed 20% of the global market
  • Transition from "following" to "running alongside" to "leading"

References​

  1. Huawei Ascend 910C — Baidu Baike
  2. Huawei Ascend series AI chip detailed parameter comparison (2025-2028) — EET-China
  3. Huawei Ascend 910C compute cluster powers domestic chip's successful trillion-scale AI large-model training — QQ News
  4. Huawei Ascend 910C completes DeepSeek V4 Pro training — Huxiu
  5. Huawei Ascend 910C measured efficiency surpasses H100, AI Infra software-hardware co-optimization shines at ten-thousand-card cluster — CNBlogs

Last updated: June 10, 2026

Computex 2026 Wrap-Up: AI PC Chip War Begins, NVIDIA RTX Spark Arrives Fall 2026

· 3 min read
Industry Research Team

June 6, 2026 — COMPUTEX 2026 concluded yesterday in Taipei. Under the theme "AI Together," this year's event set records with 1,500+ exhibitors and 6,000 booths. The head-to-head battle between NVIDIA, Intel, and AMD in the AI PC space was the defining story of the show.

1. NVIDIA RTX Spark: June Launch at $1,399​

Less than a week after its COMPUTEX debut, the NVIDIA-MediaTek RTX Spark Superchip confirmed its commercial timeline:

DetailInfo
Launch OEMsASUS, Dell, HP, Lenovo, Microsoft Surface, MSI
AvailabilityFall 2026
Starting PriceNot yet announced (analysts estimate $3,000-4,000)
Core SpecsArm CPU (up to 20 cores) + Blackwell GPU (6,144 CUDA cores)
Unified Memory128 GB LPDDR5X (300 GB/s)
Model CapacityRuns 120B parameter models, up to 1M token context

Market Reaction: AMD, Intel, and Qualcomm shares fell following the announcement. Analysts believe RTX Spark will reshape the market across three fronts — Windows AI PCs, creator workstations, and edge inference nodes.


2. Intel 18A in Full Production: Clearwater Forest + Crescent Island​

Intel CEO Lip-Bu Tan delivered his first COMPUTEX keynote with two key updates:

Clearwater Forest (Xeon 6+)​

  • 288 cores, Darkmont architecture
  • First Intel 18A process node data center CPU
  • Foveros Direct 3D packaging
  • Now in full production

Crescent Island AI GPU​

  • 480 GB LPDDR5x memory
  • 350 W air-cooled PCIe form factor
  • Native FP4 support, targeting agentic inference
  • Shipping H2 2026

"As AI moves into the agentic era, the CPU returns to the center of modern AI infrastructure." — Lip-Bu Tan


3. AMD Ryzen AI 400 Series Now Shipping​

AMD showcased the Ryzen AI 400 series (Zen 5 + Zen 5C hybrid + XDNA2 NPU) at COMPUTEX:

  • NPU performance: 60 TOPS, the highest in x86
  • 7 consumer SKUs + commercial PRO series
  • Multiple OEM models already available or launching soon
  • Advancing AI 2026 summit set for July in San Francisco

4. Chinese Domestic Chips Gaining Momentum​

VendorProductStatus
HuaweiAscend 950PR/950DTIn production, self-developed HBM
CambriconMLU6902 PFLOPS FP8, shipping
Moore ThreadsMTT S50001,000 TFLOPS, specs public

5. The AI PC Era: Three-Way Roadmap Comparison​

DimensionNVIDIA RTX SparkIntel Clearwater Forest + Crescent IslandAMD Ryzen AI 400
CPU Cores20-core Grace (Arm)288-core Darkmont (x86)Up to 12-core Zen5+5C
GPU/NPUBlackwell GPUCrescent Island (discrete GPU)XDNA2 NPU (60 TOPS)
AI Compute1 PFLOPSTBD60 TOPS NPU
TargetPersonal AI agentsDual-track: DC + AI PCCopilot+ PC
ProcessTSMC 4NPIntel 18ATSMC 4nm
AvailabilityJune 2026H2 2026Shipping now

This Week in AI Compute (6/1 – 6/6)​

DateEvent
Jun 1NVIDIA GTC Taipei: RTX Spark, Vera Rubin production, DGX Station for Windows
Jun 1Intel unveils Crescent Island, Clearwater Forest
Jun 2COMPUTEX 2026 opens: "AI Together"
Jun 5COMPUTEX closes: 1,500+ exhibitors, record scale
Jun 6RTX Spark confirmed June launch at $1,399

Sources: COMPUTEX Daily, Tencent News, Phoenix Technology, Xueqiu, The Silicon Review.

Computex 2026 AI Compute Card Major Events: DGX Station for Windows, Intel Crescent Island, and More Major Launches

· 4 min read
Industry Research Team

June 1-5, 2026, Taipei — Computex 2026 (Taipei International Information Technology Show) wrapped up successfully this week. With the theme "AI Together," industry giants including NVIDIA, Intel, AMD, and Qualcomm unveiled numerous AI compute products in rapid succession. Below, MirrorFrog brings you a roundup of the most noteworthy developments in the compute card space this week.

① NVIDIA DGX Station for Windows: A Desktop AI Supercomputer​

NVIDIA officially launched the DGX Station for Windows during its Computex 2026 keynote, calling it "the world's most powerful desktop AI supercomputer."

Core Specifications​

ItemSpecification
ChipGB300 Grace Blackwell Ultra Desktop Superchip
GPU Memory252 GB HBM3e (7.1 TB/s)
CPU Memory496 GB LPDDR5X (396 GB/s)
Unified Memory748 GB (NVLink-C2C interconnect)
FP4 Compute20 PFLOPS (sparse)
FP8 Compute10 PFLOPS (sparse)
NetworkConnectX-8 SuperNIC, up to 800 Gb/s
Model CapacityCan run 1 trillion parameter models
System Power1,600 W
Operating SystemMicrosoft Windows
ShippingQ4 2026

Significance: DGX Station compresses AI compute power (20 PFLOPS FP4) that previously required datacenter-class clusters into a single desktop workstation. 748GB of unified memory means developers can run models with hundreds of billions or even trillions of parameters locally, without cloud dependency.


② Intel Crescent Island: Inference-Specialized AI GPU​

At Computex, Intel disclosed detailed specifications for its next-generation datacenter AI inference GPU, Crescent Island.

ItemSpecification
MemoryUp to 480 GB LPDDR5x
Power350 W (PCIe form factor)
Precision SupportFP4/MXFP4 → FP64 (full precision coverage)
TargetAI inference workloads (Agentic Inference)
PositioningBetter price-performance than HBM solutions
ShippingH2 2026

Significance: Crescent Island represents Intel's key strategic move in the AI inference market. 480GB of massive LPDDR5x memory (non-HBM) means significantly lower cost compared to NVIDIA H200/B200 and other competing products, targeting enterprise inference deployment scenarios.


③ Intel Xeon 6+ (Clearwater Forest): First Intel 18A Datacenter CPU​

Intel also unveiled the new Xeon 6+ processor, codenamed Clearwater Forest, its first datacenter CPU built on the 18A process:

  • 288 Darkmont architecture cores
  • L2 288MB + L3 576MB cache
  • 12-channel DDR5-8000 memory
  • Foveros Direct 3D advanced packaging
  • AI Agent Era: CPU returns to the center of infrastructure

④ NVIDIA RTX Spark Ecosystem Takes Shape​

This week, the RTX Spark super chip developed in collaboration between NVIDIA and MediaTek continued to generate buzz. Multiple OEMs showcased RTX Spark-based laptop and compact desktop prototypes:

  • ASUS, Dell, HP, Lenovo, Microsoft Surface, MSI all confirmed as launch partners
  • Equipped with 20-core Grace CPU + Blackwell GPU (6144 CUDA cores)
  • AI compute 1 PFLOPS
  • Retail availability Fall 2026

⑤ Intel × Foxconn AI Infrastructure Partnership​

Intel and Foxconn announced a joint AI infrastructure initiative, covering the complete chain from chip → server → rack-scale system, targeting the datacenter market opportunity driven by surging AI inference demand.


⑥ Domestic AI Chip Developments​

According to the IDC 2025 annual report, total AI accelerator card shipments in China reached approximately 4 million units, with domestic vendors shipping approximately 1.65 million units, capturing a market share exceeding 41%. Huawei's Ascend 950 series has entered mass production and delivery, while Cambricon's MLU690 has begun shipping to internet customers.


This Week's Compute Roundup​

VendorProductHighlightTimeline
NVIDIADGX Station for Windows20 PFLOPS, 748GB unified memoryQ4 2026
NVIDIARTX Spark1 PFLOPS AI PC chipFall 2026
IntelCrescent Island GPU480GB LPDDR5x, 350WH2 2026
IntelXeon 6+ (Clearwater Forest)288 cores, Intel 18AH2 2026
Intel + FoxconnAI infrastructure partnershipChip→rack full chainStrategic partnership
HuaweiAscend 950PR/DT1 PFLOPS FP8, self-developed HBMIn mass production
CambriconMLU6902 PFLOPS FP8, 192GB HBM3EShipping

Sources: NVIDIA GTC Taipei 2026 / Computex 2026 official announcements, Intel press releases, ifeng Tech, IT Home.

Huawei Ascend 950 Mass Production and the Full Picture of China's AI Chip Ecosystem

· 4 min read
Industry Research Team

June 2026 — Huawei's Ascend 950 series (950PR / 950DT) has entered formal mass production and delivery, a landmark event for China's AI chip industry in 2026. Meanwhile, Cambricon's MLU690 has begun shipping and Moore Threads has announced MTT S5000 specifications, formally establishing China's tri-polar AI chip landscape.

Ascend 950 Series: A Historic Breakthrough with Self-Developed HBM​

Huawei HiSilicon's Ascend 950 series is the fourth-generation Ascend AI chip, first revealed at Huawei Connect 2025 in September and entering mass production in Q1 2026.

950PR (Prefill Inference Specialized)​

ItemSpecification
ArchitectureDa Vinci v5 (SIMD + SIMT dual-model)
ProcessN+2 (SMIC domestic)
HBMHiBL 1.0 (Huawei self-developed) , 128 GB
FP8 Compute1 PFLOPS (HiF8 format)
TDP~400 W
TargetInference Prefill (video recommendation, real-time interaction)

950DT (Decode + Training Specialized)​

ItemSpecification
ArchitectureDa Vinci v5 (SIMD + SIMT dual-model)
ProcessN+2 (SMIC domestic)
HBMHiZQ 2.0 (Huawei self-developed) , 144 GB, 4 TB/s
FP8 Compute1 PFLOPS (HiF8 format)
TDP~500 W
TargetInference Decode + Model Training

Historical Significance​

Self-developed HBM (HiBL 1.0 / HiZQ 2.0) represents the most important technical breakthrough of Huawei Ascend 950 — this is the first time a Chinese enterprise has achieved self-developed mass production of HBM memory, completely eliminating dependence on SK Hynix / Samsung HBM supply. Combined with the domestic N+2 process, Ascend 950 has achieved full-chain domestic production from HBM → Compute Die → Packaging → System.

Cambricon MLU690: China's Only Native FP8 Support​

Cambricon's seventh-generation AI chip MLU 690 (Siyuan 690) began volume production and shipping in H1 2026. This is the first domestic AI chip with native FP8 precision support.

ItemMLU 690
Process5nm (TSMC / SMIC)
FP8 dense2 PFLOPS
HBM192GB HBM3E, 5 TB/s
TDP~500 W
Unit Price (OAM)~$8,000-12,000

MLU 690's FP8 compute power (2 PFLOPS dense) is on paper comparable to NVIDIA Blackwell (B200 FP8 4.5 PFLOPS sparse). Leveraging its financing advantage as a STAR Market listed company, Cambricon targets 2026 revenue of ¥15-20B (2025: ¥7.2B).

Moore Threads MTT S5000: From Graphics to Training-Inference Unified​

Moore Threads publicly disclosed detailed specifications of the MTT S5000 in February 2026, featuring the fourth-generation MUSA "Pinghu" architecture, single-card AI compute of 1,000 TFLOPS, 80GB GDDR6X memory, 1.6 TB/s bandwidth.

Moore Threads pursues a full-function GPU path (graphics rendering + AI compute + general-purpose compute), closest to NVIDIA's strategy. The founding team comes from former NVIDIA China, and the MUSIFY toolchain helps auto-migrate CUDA code to the MUSA platform, lowering ecosystem migration costs.

China's Tri-Polar AI Chip Landscape​

DimensionHuawei AscendCambriconMoore Threads
Core ArchitectureDa Vinci v5MLUv07MUSA 4th Gen
ProcessN+2 domestic5nm6nm
FP8 Compute~1 PFLOPS2 PFLOPS0.5 PFLOPS (estimated)
HBM Self-Sufficiency✅ Self-developed HiBL/HiZQ❌ Purchased❌ Purchased
EcosystemCANN + MindSporeNeuWare + MindSporeMUSA + MUSIFY
AdvantageFull-chain domesticHighest FP8 computeFull-function + CUDA migration
2025 Revenue(Huawei internal)¥7.2B¥2.2B

Global Market Comparison (Q2 2026 Update)​

TierVendorFlagship ChipFP8/PFLOPSHBMMass Production
Tier 1NVIDIARubin R20025 PF (sparse)288GB HBM42026 H2
Tier 2AMDMI40020 PF (dense)432GB HBM42026
HuaweiAscend 950DT1 PF (dense)144GB self-developed HBM2026 Q1
CambriconMLU6902 PF (dense)192GB HBM3E2026 H1
AWSTrainium 35.7 PF (dense)144GB HBM2025 Q4 GA
Tier 3IntelGaudi 31.8 PF128GB HBM2eIn production
GoogleTPU v74.6 PF(TFLOPS)192GB HBM2025
Moore ThreadsMTT S50001 PF80GB GDDR6X2025 Q1

Note: NVIDIA uses sparse compute as standard, while AMD / Huawei / Cambricon use dense — not directly comparable.

Outlook for H2 2026​

  • NVIDIA Rubin R200: Official shipment in H2 2026, 288GB HBM4, 6-chip CoWoS-L packaging
  • Huawei Ascend 960: Roadmap H2 2027, expected FP8 compute doubled to 2 PFLOPS
  • Cambricon MLU790: Expected 2027, 3nm, 384GB HBM4, 2.5 PFLOPS
  • Moore Threads: Next-gen GPU expected with HBM3, 2× MTT S5000 compute

By 2026, China's AI chip industry has formed a complete product matrix from Training (Cambricon MLU690 / Ascend 950DT) → Inference (Ascend 950PR / Moore Threads S5000) → Systems (CloudMatrix / Distributed Clusters).


This article is based on public information from Huawei Connect 2025 (2025-09-18), industry analysis reports from April 2026, and the latest market data as of June 2026.