AI Accelerators

Twitter

Facebook

Copy Link

Key Facts

Key ChipsH100, MI300X, Gaudi 3

Key PlayersNVIDIA, AMD, Intel

Top News

Google Teams with AMD on 10th-Generation TPU Project for Reinforcement Learning Workloads

Google is reportedly collaborating with AMD to develop its 10th-generation TPU, marking AMD's first major involvement in a custom AI accelerator project. The hybrid design may integrate on-package CPU cores alongside tensor compute chiplets to handle reinforcement learning and agentic AI workloads that demand more general-purpose processing power.

Technology2 sources

Google Teams with AMD on 10th-Generation TPU Project for Reinforcement Learning Workloads

Technology2 sources

Cerebras Systems Launches CS-4 Rack Systems to Supercharge AI Hardware Performance

Cerebras Systems unveiled its CS-4 rack powered by WSE-3T chips, delivering 10x throughput per watt over previous generation. The AI hardware now powers OpenAI's GPT-5.6 Sol Ultrafast mode at 750 tokens per second. With 21.6 petabytes per second memory bandwidth, Cerebras targets faster AI inference to speed AI chatbots and compete with Nvidia.

Technology7 sources

Cerebras Systems Launches CS-4 Rack Systems to Supercharge AI Hardware Performance

Technology7 sources

Nvidia RTX Pro 6000 GPU Price Nearly Doubles to $16,000 Amid AI Hardware Demand

Nvidia's flagship RTX Pro 6000 workstation GPU has surged to $16,000, nearly double its March 2025 launch price of $8,565. The AI-focused Blackwell GPU features 96GB of GDDR7 memory and 24,064 CUDA cores, but soaring component costs and memory shortages are driving unprecedented price increases across professional graphics cards.

Business2 sources

Nvidia RTX Pro 6000 GPU Price Nearly Doubles to $16,000 Amid AI Hardware Demand

Business2 sources

Modified GPUs with doubled VRAM surface on eBay and Chinese markets as AI hobbyists chase memory

Modified GPUs are flooding second-hand markets as AI enthusiasts seek affordable VRAM capacity. A Hong Kong-based seller is offering modded RTX 2080 Ti cards with 22GB VRAM for $499 on eBay, while modified RTX 4080 graphics cards with 32GB are appearing in China's Goofish marketplace for approximately $1,479. These unofficial upgrades double original memory pools without manufacturer warranties but offer budget-friendly paths to running large language models and diffusion models locally.

Technology6 sources

Modified GPUs with doubled VRAM surface on eBay and Chinese markets as AI hobbyists chase memory

Technology6 sources

AMD Acquires Taalas to Etch AI Models Into Silicon for Faster Inference

AMD acquired Toronto-based Taalas, an AI chip startup that hardwires model weights directly into silicon. The deal strengthens AMD's AI inference capabilities with technology that promises 48x faster performance than traditional GPUs, though chips remain locked to specific models.

Technology11 sources

AMD Acquires Taalas to Etch AI Models Into Silicon for Faster Inference

Technology11 sources

NVIDIA Open-Sources cuFile API and Launches Storage-Next Initiative With 40+ Industry Partners

NVIDIA announced it is open-sourcing its cuFile API at the Future of Memory and Storage conference, enabling direct GPU-to-storage data access in microseconds. The company also launched Storage-Next, bringing together over 40 storage vendors including DDN, Kioxia, and Micron to standardize GPU-driven storage solutions for AI workloads.

Technology4 sources

NVIDIA Open-Sources cuFile API and Launches Storage-Next Initiative With 40+ Industry Partners

Technology4 sources

Majestic Labs unveils Prometheus AI server with 128TB memory to challenge Nvidia's dominance

Tel Aviv startup Majestic Labs has unveiled Prometheus, a GPU-free server that uses up to 128TB of LPDDR6 memory and Arm cores to tackle AI's memory wall. The company claims one rack matches 25 Nvidia Vera Rubin racks at lower power, but hardware hasn't shipped or been independently tested yet.

Technology2 sources

Majestic Labs unveils Prometheus AI server with 128TB memory to challenge Nvidia's dominance

Technology2 sources

Eliyan raises $145M at $1B valuation to tackle AI chip data bottlenecks in data centers

Silicon Valley startup Eliyan has raised $145 million in Series C funding at a $1 billion valuation, backed by Cisco and Lumentum. The company develops electro-optical interconnects to solve a critical problem: AI chips can process data much faster than they can send or receive it, leaving expensive GPUs sitting idle 60-70% of the time.

Startups2 sources

Eliyan raises $145M at $1B valuation to tackle AI chip data bottlenecks in data centers

Startups2 sources

AMD and Cerebras forge strategic partnership to challenge Nvidia's $20 billion Groq acquisition

AMD and Cerebras Systems announced a strategic partnership to develop a disaggregated AI inference platform that combines AMD's Helios rack-scale infrastructure with Cerebras' Wafer-Scale Engine. The collaboration promises up to 5x higher tokens per second per watt by splitting inference workloads between AMD's EPYC processors and Instinct GPUs for prompt processing, while Cerebras handles memory-intensive token generation with its SRAM-powered accelerators.

Technology10 sources

AMD and Cerebras forge strategic partnership to challenge Nvidia's $20 billion Groq acquisition

Technology10 sources

Geekbench 7 launches with overhauled tests for AI workloads, CUDA support, and real-world benchmarking

Primate Labs released Geekbench 7, introducing its most significant update yet with overhauled performance tests for CPU and GPU workloads. The cross-platform benchmark tool now features AI benchmarks, CUDA support for Nvidia GPUs, and real-world multi-core benchmarking that mirrors actual application behavior. New media workloads include AV1 encoding, Opus audio compression, and Whisper speech recognition for live captions.

Technology2 sources

Geekbench 7 launches with overhauled tests for AI workloads, CUDA support, and real-world benchmarking

Technology2 sources

Weka launches NeuralMesh 6 to cache 100% of tokens and solve GPU memory bottlenecks in AI inference

Weka unveiled NeuralMesh 6, its biggest software overhaul, alongside its first custom-designed hardware line WEKApod. The platform introduces augmented memory grid technology that caches 100% of pre-calculated tokens, preventing GPUs from wasting compute on redundant calculations. Benchmark tests show 10x higher token throughput and 10x more concurrent users compared to standard memory systems.

Technology2 sources

Weka launches NeuralMesh 6 to cache 100% of tokens and solve GPU memory bottlenecks in AI inference

Technology2 sources

Snapdragon 4 Gen 6 leak reveals on-device AI headed to budget Android phones under $400

A leak suggests Qualcomm's upcoming Snapdragon 4 Gen 6 will bring dedicated AI processing to budget Android phones for the first time. The chip features a Hexagon DSP and separate machine learning engine, potentially enabling live translation, smarter camera tools, and voice features on entry-level devices costing under $400.

Technology2 sources

Snapdragon 4 Gen 6 leak reveals on-device AI headed to budget Android phones under $400

Technology2 sources

Google develops Frozen v2 chip with Gemini architecture hardwired into silicon for efficiency

Google is building a specialized AI chip called Frozen v2 that embeds Gemini's architecture directly into hardware. The chip could deliver 6-10 times better efficiency than current TPUs, addressing severe compute shortages that forced Google Cloud to turn away customers. Deployment targets 2028, but the approach trades flexibility for performance.

Technology7 sources

Google develops Frozen v2 chip with Gemini architecture hardwired into silicon for efficiency

Technology7 sources

Alphabet's Frozen v2 chip could deliver 10x efficiency gains by embedding Gemini into silicon

Alphabet is developing Frozen v2, a server chip that embeds its Gemini AI model directly into hardware to achieve 6-10 times more tokens per unit of power than current TPUs. The move signals a strategic shift from spending more on AI infrastructure to making AI operations dramatically cheaper, with deployment targeted for 2028.

Technology3 sources

Alphabet's Frozen v2 chip could deliver 10x efficiency gains by embedding Gemini into silicon

Technology3 sources

Cadence AuraStack AI agent cuts design time in half, delivers 15X productivity boost for engineers

Cadence Design Systems launched AuraStack, an agentic AI platform that automates PCB and advanced packaging design workflows. The AI agent delivers up to 2X faster time to market and 15X productivity gains by orchestrating complex engineering tasks through natural language commands. Nvidia, TSMC, and Schneider Electric are already deploying the system to tackle bottlenecks that consume 65% of engineers' time.

Technology6 sources

Cadence AuraStack AI agent cuts design time in half, delivers 15X productivity boost for engineers

Technology6 sources

Prev12 3 4…7 Next