AI Chips & Infrastructure
Silicon accelerators, TPU clusters, custom ASICs, optical interconnects, and datacenter energy innovations.
Published Reporting (5 Stories)
Sorted by recency
NVIDIA Blackwell B200 Superchips Reach Volume Datacenter Deployment Across Hyperscalers
NVIDIA has confirmed the commercial deployment of GB200 NVL72 liquid-cooled racks, providing up to 30x inference speedup for trillion-parameter generative models while slashing energy costs by 25x.
Cerebras CS-3 Wafer-Scale Supercomputer Sets Record 125 Petaflops AI Compute on a Single Silicon Wafer
Cerebras Systems has demonstrated the CS-3 system, a single-wafer processor containing 4 million AI-optimized cores and 44GB of on-chip SRAM, eliminating interconnect latency bottlenecks for large model training.
Groq LPUs Reach 500 Tokens Per Second Inference Speeds with Deterministic Tensor Streaming
Groq Language Processing Units (LPUs) demonstrate deterministic 500+ token/second generation rates for open-weights models, creating new possibilities for real-time conversational voice agents.
AMD Instinct MI325X Accelerators Feature 256GB HBM3E Memory for Trillion-Parameter Model Training
AMD has officially launched the Instinct MI325X GPU accelerator, delivering 256GB of ultra-fast HBM3E memory and 6.0 TB/s bandwidth to challenge NVIDIA datacenter dominance.
AWS Trainium2 Clusters Scale to 100,000 Custom Silicon Nodes for Hyperscale Foundation Models
Amazon Web Services has deployed UltraClusters of 100,000 Trainium2 chips, providing 65 Exaflops of aggregate compute for training next-generation frontier AI models.