Silicon for GPU/XPU

 

Huawei Ascend

• Huawei announces annual release cadence for three new Ascend AI chips, unveils ‘supernode’ offering company says will outperform Nvidia’s NVL144
• "The roadmap was unveiled at the company’s Huawei Connect event taking place in Shanghai this week. The Chinese tech giant said it would release its Ascend 950 chip – the successor to its 910C AI chip launched earlier this year – in 2026, with the Ascend 960 and Ascend 970 to follow in 2027 and 2028, respectively." [DCD-20250925]

Cambricon Technologies

• Siyuan 690
• "Shares of his chip designer Cambricon Technologies have surged more than 765% over the past 24 months."[The Edge - 20251118]

Chiplet Interconnect

• Making connections: The pursuit of chiplet interconnect standardization
• "A closer look at Universal Chiplet Interconnect Express technology" [DCD - 20250918]

Groq

• AI chip company Groq raises $750m at $6.9bn valuation
• "Round was led by investment firm Disruptive" [DCD - 20250918]

Tesla

• Tesla's Chip Bet: Beyond Cars, Beyond Vision, What's the Ultimate Goal?
• On September 7, Tesla CEO Elon Musk stated on social media platform X, "Just had an excellent design review with the Tesla AI 5 chip design team." He described this chip as an "epic" product and revealed that the development of its successor, the AI6 chip, has been scheduled. [MM-20250910]

ARM

• How Arm is building infrastructure for a systems-level world
• "Chip design giant has gone from playing catch-up to being a market-leading triple threat in 15 years". [DCD-20250909]

AWS

• Graviton4 
• Graviton4, it has 30% better performance than Graviton3, while also containing 50% more cores and up to 75% better memory. [AWS-20231129]
• Trainium2
• Trainium2 is capable of delivering up to four times faster training than the first generation and will be used in AWS's EC2 UltraClusters of up to 100,000 chips. The cloud giant said it can train foundation and large language models, or LLMs, in "a fraction of the time" and improve energy efficiency by up to two times.
• Amazon (AMZN) said Anthropic, Databricks, Datadog, Epic, Honeycomb and SAP are among AWS customers using the new chips. 
• Further references, SeekingAlpha.

Microsoft

• Maia 100
• "The Maia 100 chip is aimed at AI workloads, and could go up against offerings from Nvidia (NVDA). It will be available for Azure cloud customers and it is already being tested with Bing and Office AI products, Microsoft executive Rani Borkar said at the conference, according to Bloomberg." [SA-20241115]
• Cobalt 100
• "The Cobalt 100 chip, which uses Arm Holding's (ARM) architecture, is aimed at general computing and could compete with offerings from Intel (INTC) and AMD (AMD) in the data center space."

Google

• TPUv7: Google Takes a Swing at the King
• "Potential End of the CUDA Moat?, Anthropic’s 1GW+ TPU Purchase, The more (TPU) Meta/SSI/xAI/OAI/Anthro buy the more (GPU capex) you save, Next Generation TPUv8AX and TPUv8X versus Vera Rubin" [SA-20251128]
• Google AI Infrastructure Supremacy: Systems Matter More Than Microarchitecture
• "From DLRM to LLM, internal workloads win, but how does Google fare in external workloads" [SA-20230423]
• Google says TPU demand is outstripping supply, claims 8yr old hardware iterations have “100% utilization”
• "The future of processors lies in specialized architecture, Google’s VP and GM of AI and infrastructure, Amin Vahdat, said" [DCD-20251030]
• TPU v5e
• "Cloud TPU v5e is purpose-built to bring the cost-efficiency and performance required for medium- and large-scale training and inference. TPU v5e delivers up to 2x higher training performance per dollar and up to 2.5x inference performance per dollar for LLMs and gen AI models compared to Cloud TPU v4. At less than half the cost of TPU v4, TPU v5e makes it possible for more organizations to train and deploy larger, more complex AI models. " [Google-20230830]
• TPU v6 
• "Google makes 6th generation TPUs generally available" [Google-20241212]
• "The TPUs were used to train its Gemini 2.0 AI model, and are a key component of Google Cloud's AI Hypercomputer."
• An in-depth look at Google’s first Tensor Processing Unit (TPU) 
• "There’s a common thread that connects Google services such as Google Search, Street View, Google Photos and Google Translate: they all use Google’s Tensor Processing Unit, or TPU, to accelerate their neural network computations behind the scenes" [Google - 20170513]
• 
  
  
  

Meta

MTIA (V2)

• "The next generation of MTIA is part of our broader full-stack development program for custom, domain-specific silicon that addresses our unique workloads and systems. This new version of MTIA more than doubles the compute and memory bandwidth of our previous solution while maintaining our close tie-in to our workloads. It is designed to efficiently serve the ranking and recommendation models that provide high-quality recommendations to users." [META-20240410]

• MTIA (V1)
• "We found that GPUs were not always optimal for running Meta’s specific recommendation workloads at the levels of efficiency required at our scale. Our solution to this challenge was to design a family of recommendation-specific Meta Training and Inference Accelerator (MTIA) ASICs. We co-designed the first-generation ASIC with next-generation recommendation model requirements in mind and integrated it into PyTorch to create a wholly optimized ranking system. In addition, we maintained the user experience and developer efficiency offered by PyTorch eager-mode development. Developer efficiency is a journey as we continue to support PyTorch 2.0, which supercharges how PyTorch operates at the compiler level — under the hood.." [META-20230518]

Intel

• Crescent Island
• "Power and cost-optimized for air-cooled enterprise servers" [DCD - 20251015]
• Jaguar Shores
• "Intel has put the name of a brand new AI chip on its roadmap that will compete with offerings from AMD and Nvidia. The Jaguar Shores AI chip is listed as a successor to Falcon Shores, a GPU for AI that is due for release next year.." [HPC-20241119]
• Falcon Shores
• "Intel is planning to onboard a new version of the Falcon Shores chip in 2026, which is code-named Falcon Shores 2. " [HPC-20230808]