[BidClub_]
SemiAnalysis · · 78 分钟

AI芯片与硅产业盘点 2026

Jordan NanosRay WangMalcolm SplitterJoey Brookhart

YouTube
TL;DR
  • NVIDIA的Vera Rubin VR200被称为2026年最受期待的发布,叙述者预计它“很可能再次登顶”。 该芯片基于TSMC N3B工艺,配备HBM4,单个封装目标是35 PFLOPS FP4算力,拥有288 GB显存、22 TB/s带宽,并通过NVIDIA的NVLink扩展网络扩展成NVL72机架——但需要注意,22 TB/s只是“NVIDIA目前设定的目标速度”。
  • AMD的2026年可能成为转折点,前提是执行到位:“如果MI455X和Helios Rack能按时交付”。 CDNA 5采用3200亿个晶体管,由12个2纳米和3纳米逻辑芯粒组成,并通过先进的3.5D封装连接;其核心优势在于内存:432 GB新一代HBM4,带宽接近20 TB/s。这一优势将“至少持续到Rubin Ultra发布,后者将配备完整的1 TB HBM”。
  • 本期的结构性判断是:“越来越多的推理任务正在从NVIDIA转向定制芯片。” Microsoft Maia 200拥有1400亿个晶体管、216 GB HBM3E,在FP8和FP4下的算力分别超过5 PFLOPS和10 PFLOPS,未来将运行 ChatGPT 模型;Meta的MTIA v3很可能采用HBM,并且“能带来不错的利润率”,因此外部硬件可以承担训练,内部推理则转向自研。
  • Trainium 3凭借Trainium 2的部署规模,被视为未来大规模部署的头号竞争者。 AWS在Canton和New Carlisle数据中心已经部署了数十万颗Trainium 2;Anthropic的Claude Code在Trainium上训练并运行,而“OpenAI将从今年开始使用2 GW的Trainium算力”。
  • 如果外部客户能像Google一样使用Ironwood(TPU v7),它将成为效率领域的重量级产品。 该芯片配备192 GB HBM3E,很可能拥有超过1000亿个晶体管,并通过光学电路交换机——“微小的物理镜面”——连接最多容纳9,216颗TPU的superpod。
  • Qualcomm押注LPDDR,1年前可能是个绝妙主意;但如今内存价格“全面飙升”,LPDDR5X也不例外。 搭载768 GB LPDDR5X的AI 200不太可能掀起波澜,但AI 250据称将采用计算贴近内存的架构;Qualcomm称,搭配LPDDR6后,该架构能带来10倍有效带宽提升——“先别急着戴上派对帽,但要记住Qualcomm”。
  • SRAM路线的异类产品值得尊重,但还不足以形成确信:NVIDIA认为Groq的确定性LPU“价值大约200亿美元”,而Cerebras搭载44 GB片上SRAM的WSE3“已经不太够用了”,即便它采用的是SRAM。 Intel的Jaguar Shores采用18A工艺并配备288 GB HBM4,纸面上具备竞争力,但目前看并非面向2026年,可能要到2027年才推出。
摘要 · 为研究而整理的核心内容

参考音频是一段由单一叙述者完成的脚本式盘点,以下引述均归于该叙述者。叙述者将2026年定义为AI硬件的大年,覆盖GPU、ASIC及更多方向,并指出这个行业已经不再处于萌芽期:Google在2015年推出TPU,NVIDIA则在2017年推出首款Tensor Core GPU。

1. Vera Rubin与Helios——内存容量成为新战场

  • 叙述者对NVIDIA的判断是:采用TSMC N3B工艺、配备HBM4的VR200,单个封装提供35 PFLOPS FP4算力、288 GB容量和目标22 TB/s带宽;放进NVL72机架后,“很可能再次登顶。还是说,不会?”
  • 真正需要关注的变量是:MI455X拥有3200亿个晶体管、432 GB新一代HBM4和约20 TB/s带宽,在内存上占优,“至少持续到Rubin Ultra发布,后者将配备完整的1 TB HBM”。如果MI455X和Helios Rack按时交付,AMD的2026年可能成为转折点。
  • Intel的Jaguar Shores采用18A工艺、1750亿个晶体管和288 GB HBM4,纸面上具备竞争力,但目前看并非面向2026年,可能要到2027年才推出。Intel必须证明芯片本身和配套软件都能过关。

2. 超大规模云厂商的ASIC正在把推理任务从NVIDIA手中夺走

  • 核心脉络是:“越来越多的推理任务正在从NVIDIA转向定制芯片。” Maia 200芯片面积约825 mm²,拥有1400亿个晶体管、216 GB HBM3E,在FP8和FP4下的算力分别超过5 PFLOPS和10 PFLOPS,未来将运行 ChatGPT 模型。
  • 按叙述者的说法,Meta“非常清楚自己的内部工作负载需要什么”——包括AI聊天机器人和推荐模型,但不是 AGI——因此自研推理芯片可以提供不错的利润率,同时让Meta用外部硬件训练新模型。具体规格尚未公开,但这款芯片很可能采用TSMC N3P工艺,晶体管数量超过1000亿,并配备HBM;此前的MTIA版本采用LPDDR5X。
  • Trainium 3是未来大规模部署的头号竞争者:它同时支持训练和推理,而AWS的Canton和New Carlisle数据中心已经部署了数十万颗Trainium 2。Claude Code在Trainium上训练并运行,OpenAI则将从今年开始使用2 GW的Trainium算力。

3. Google凭借光交换技术享有十年先发优势

  • Ironwood(TPU v7)采用N3E工艺,很可能拥有超过1000亿个晶体管,并配备192 GB HBM3E,主要用于运行Gemini推理任务;它真正的“超能力”是光学电路交换机,也就是连接最多容纳9,216颗TPU的superpod的物理镜面。
  • 到目前为止,TPU只有Google自己在使用;只有当外部客户也能像Google一样使用Ironwood时,它才会成为“高效、低成本推理领域的重量级产品”。

4. Qualcomm的内存押注与SRAM路线的逆行者

  • Qualcomm近期搭建的1,024颗芯片AI 100集群规模很小,“对2026年无关紧要”。AI 200选择LPDDR5X而不是供应受限的HBM,这个押注1年前可能很漂亮,但如今内存价格“全面飙升”,LPDDR5X也未能幸免。AI 200今年不太可能掀起大波澜,市场希望转向AI 250的计算贴近内存设计:Qualcomm宣称,搭配速度更快的新一代LPDDR6后,有效带宽可提升10倍。
  • Groq——“名字里是Q,不是K”——用内存换取确定性:230 MB SRAM、不依赖外部内存,并通过类似钟表般的执行机制,“基本消除了GPU需要处理的延迟问题”;但“即便运行更小的模型,也需要连接大量LPU。一切都是权衡。”NVIDIA收购了Groq,并认为其价值约200亿美元。2026年是否会有产品尚不确定,真正值得关注的是基于Samsung 4纳米工艺的第二代LPU。
  • Cerebras的晶圆级WSE3拥有4万亿个晶体管、44 GB SRAM,理论带宽为21 PB/s,在超高速服务器领域依然“非常独特”且具备竞争力;但“即便是SRAM,44 GB也已经不太够用了”。目前只能把希望寄托在WSE4官宣上。

5. 尚未回答的宏观问题

  • 盘点结束时,叙述者留下了一个悬而未决的问题:“现在,所有曲线都还在向上,但这种情况还能持续多久?”随后话题转向SemiAnalysis的Accelerator和HBM模型,以及覆盖真实加速器基准测试的免费Inference X仪表盘。
Speaker 1

2026 will be a massive year for AI hardware, no matter whether we're talking about GPUs, ASICs, or beyond. I'm not sure we can still call this an emerging industry. Google's first TPU was released in 2015, over a decade ago, and NVIDIA will celebrate its 10-year AI anniversary next year because Volta, the first Tensor Core GPU, launched in 2017. So it's settled: this is not a new industry anymore, but there are plenty of new chips. Here are, in no particular order, the most interesting AI chips for 2026. Let us know which chip you think will come out on top, and leave a comment if we missed a chip you think deserves to be in the next video.

1. Qualcomm Targets Inference

Let's start with a company and a chip whose first major deployment was recently announced. What was deployed, you ask? It was 3-year-old hardware based on an almost 6-year-old architecture. I'm talking about Qualcomm finally deploying its AI 100 chips at scale, even if the scale was small. Installing a cluster of 1,024 AI 100 chips doesn't matter in 2026. But what might matter is the new AI 200 chip Qualcomm announced in October last year.

With about 70 billion transistors produced on TSMC's N3E and 768 GB of LPDDR5X memory, the AI 200 ASIC is clearly designed for inference. Betting on LPDDR5X instead of supply-constrained HBM might have been a great idea 1 year ago, but memory prices are skyrocketing across the board, including LPDDR5X. It's unlikely that AI 200 will make major waves this year, but its successor, AI 250, is supposed to come with a new compute-near-memory architecture. According to Qualcomm, that results in a 10x increase in effective memory bandwidth.

Paired with fast next-generation LPDDR6 memory, there might be something worthwhile on the horizon. So don't put on your party hats just yet, but do keep Qualcomm in mind.

2. AMD Builds A Challenger

If NVIDIA is Goliath, this company would be David. At least, I'm sure that's what AMD would like to hear. AMD is the second major player in the GPU space, and 2026 could be a turning point if MI455X and the Helios Rack are on time. MI455X is based on the new CDNA 5 architecture and packs 320 billion transistors into a mix of 12 2-nanometer and 3-nanometer logic chiplets connected via advanced 3.5D packaging. But the biggest selling point could be the massive memory setup.

MI455X comes with 432 GB of next-generation HBM4 and a bandwidth of almost 20 TB/s. There's much more to MI455X and the Helios Rack than we have time for this video. If you want to know how AMD is challenging Nvidia, check out the semi-analysis article on AMD's AI strategy. It not only offers a deep dive into MI455X, but also explains the Helios Rack architecture. It's definitely worth a read. I'll drop a link in the video description below.

3. Google Extends TPU Access

Google is truly the grandfather of AI. Not only was “Attention Is All You Need” basically a Google paper, but TPUs have been kicking it since 2015. While everyone was hyped about blockchains, Google was already preparing for AI. Ironwood, aka TPUv7, is built on TSMC's N3E with very likely more than 100 billion transistors. It has, of course, 2 large compute chiplets and comes with 192 GB of HBM3E. It's specifically designed to run Gemini inference.

But Google's superpowers are the so-called optical circuit switches. These are tiny physical mirrors that allow for a fast and super-efficient optical interconnect. Google uses them to create superpods. These superpods connect up to 9,216 TPUs. My further reading recommendation is, as usual, the SemiAnalysis TPU V7 deep dive. TPUs offer a completely different approach to AI workloads and have a very different TCO metric. Link in the description below.

TPUs have always been at the forefront, but so far they were only used by Google. This changes with Ironwood. The question is whether external customers can use them as well as Google can. If so, Ironwood will be a heavyweight for efficient and low-cost inference.

4. Specialized Chips Trade Speed for Scale

Another chip that isn't really brand new but is still hot for 2026 is Cerebras' Wafer Scale Engine 3, or WSE3. It was launched back in 2024, but new clusters have just recently been announced. WSE3, as the name implies, uses an entire silicon wafer to create a massive chip. It contains an unimaginable amount of SRAM. No, not megabytes. I'm talking about 44 GB and a theoretical memory bandwidth of 21 PB/s. But remember, we are talking about an entire wafer.

Based on the slightly older N4P node, it contains a whopping 4 trillion transistors. Everything about WSE3 is extraordinary. Even if it's SRAM, 44 GB isn't really cutting it anymore. It's still a very unique concept and competitive when it comes to super-fast servers. While WSE3 could be the odd one out in 2026, we might see an announcement of WSE4. Fingers crossed.

Groq is an interesting one, not to be confused with Elon Musk's Grok AI model. The one I'm talking about has a Q, not a K. And that Groq with a Q was acquired by NVIDIA. We don't know for sure if there will be a 2026 product or if it's more of a foundation for the future, but I still wanted to mention it because it's a very cool concept.

The Groq Language Processing Unit, or LPU for short, has relatively unspectacular specs: 55 billion transistors on a 14-nanometer GlobalFoundries node and no external memory at all. Nothing to write home about, it seems. But the LPU doesn't need any of that. It has 230 MB of ultra-fast SRAM placed very close to the compute cores. The chip is designed in a way that every step, every calculation, runs like clockwork.

This is called deterministic execution, and it basically eliminates the latency problems that GPUs have to deal with. But because you have so little memory, you need to connect a lot of LPUs to run even smaller models. Everything is a trade-off. No matter what, NVIDIA thought it was worth about $20 billion. I'm excited to see what the 2nd-generation LPU based on Samsung's 4-nanometer process can do, and what NVIDIA does with it.

5. NVIDIA Raises the AI Bar

I can't avoid the elephant in the room any longer. NVIDIA is number 1 for a reason, and with Vera Rubin, NVIDIA is taking the next step, moving to TSMC's N3B and HBM4. A single VR200 can crank out an insane 35 petaflops of FP4 per package. Combined with 288 GB of ultra-fast HBM4 at 22 TB/s of bandwidth—or at least that's the speed NVIDIA is targeting—a single superchip is already a force to be reckoned with.

But combining 72 of these in the NVL72 rack using NVIDIA's NVLink scale-up network, Vera Rubin will very likely once again top the charts. Or will it? MI455X does have a memory advantage, at least until Rubin Ultra, which will come with a full terabyte of HBM. No matter what, there's no doubt that VR200 is the most anticipated release for 2026. And it's also the topic of the latest semi-analysis article, just in case you really want to know how Vera Rubin is built and how it works.

6. Hyperscalers Build Custom Silicon

Mark Zuckerberg is on a massive spending spree. Meta is basically buying everything it can, and that includes GPUs from NVIDIA and AMD, and even TPUs from Google. This makes it easy to forget that Meta has its own silicon. The Meta Training and Inference Accelerator, or MTIA for short, is now in its 3rd iteration. We don't know all the specs yet, but it will be produced on TSMC's N3P with very likely more than 100 billion transistors and HBM memory.

Previous MTIA versions still use LPDDR5X. Meta knows exactly what its internal workloads need, and those workloads aren't only AI chatbots; they're also AI recommendation models that run the algorithms for Facebook, Instagram, and Threads. MTIA V3 won't be Meta's chip for AGI, but it offers good margins for its business model.

That means Meta can use all the external hardware to train new models while internally switching to its own silicon for inference. Looking at the roadmap, we will see a lot more MTIA in the future.

There's a lot of AI silicon, but only a few chips see truly large-scale deployment. One of them is Amazon's Trainium. There are hundreds of thousands of Trainium 2 chips deployed inside AWS AI data centers in Canton and New Carlisle, which makes Trainium 3 a top contender for future large-scale deployment. Trainium 3, despite its name, combines training and inference capabilities in a single chip.

Built on TSMC's N3P, it consists of about 125 billion transistors and comes with 144 GB of fast HBM3E. You can find much more, including TCO numbers, in the semi-analysis Trainium 3 deep dive. And yes, it's an actual technical deep dive. Anthropic's amazing Claude Code was trained on and runs on Trainium. It can only get better with Trainium 3. And Anthropic isn't the only fan. OpenAI will use 2 GW of Trainium compute starting this year. If you're looking for an ASIC with truly large-scale deployment, look no further than Trainium 3.

Microsoft's first ASIC, Maia 100, was a little bit of a mystery. Like every new in-house design, it had to prove itself first. Does it warrant a 2nd generation? With Maia 200, we have the answer. At around 825 mm², the almost reticle-busting chip contains 140 billion transistors, is manufactured on TSMC's N3P, and comes with a pretty large 216 GB of HBM3E.

Maia 200 will be used for inference and is optimized for FP8 and FP4, providing over 5 and 10 petaflops, respectively. Microsoft will use it for its in-house models, but Maia 200 will also run future ChatGPT models. More and more inference is moving away from NVIDIA to custom silicon.

7. Intel Tries Again

Last but certainly not least, we have Intel, which is having another go at an AI GPU. But it doesn't seem like Jaguar Shores is targeting a 2026 release. With so many ASICs, I had to talk about one more GPU, even if it might be a 2027 product. Jaguar Shores comes with a pretty strong spec sheet: an 18A process node, 175 billion transistors, and 288 GB of HBM4.

On paper, it's competitive. But paper is patient. After many attempts, Intel not only has to prove it can build and produce an AI GPU, but also provide the proper software support. No matter what, a third player in the GPU space would be very welcome. Jaguar Shores will show whether Intel can compete again.

The chip wars are heating up, and what we just talked about is only a small selection of the entire market. If you compare this list with what the Semi Analysis Accelerator and HBM model tracks, you will know what I mean. Right now, all lines are only going up, but for how long? If you are working in or with the industry and are interested in a clear picture view of the entire AI silicon market, and where it's headed, the Semi Analysis Accelerator and HBM model is your best choice. Not only does it offer a comprehensive overview of all current and future chips, it also provides shipment numbers and ASPs. Plus, deep insights into the HBM market. You can find the link in the description below, right next to all the deep dive articles we talked about in the video. And if you are less interested in specs, but more in real-world performance, Inference X offers comprehensive benchmarks across a wide range of accelerators. Because, what really matters in the end is how a chip performs. Check out the Semi Analysis Inference X dashboard. It's free. You know where you can find the link. I hope you enjoyed this video, and see you in the next one.

AI芯片与硅产业盘点 2026 — 文字稿与摘要 | BidClub