Samsung and SK Hynix started mass production of HBM4 for NVIDIA AI accelerators https://sudonull.com/samsung-and-sk-hynix-started-mass-production-of-hbm4-for-nvidia-ai-accelerators
South Korean giants released HBM4 for NVIDIA: prices $700, supply delays, impact on AI market. Who wins and why regular DRAM is more profitable. Read the analysis.
Atlas 350: Ascend 950PR 3 times more powerful than H20 in FP4 https://sudonull.com/atlas-350-ascend-950pr-3-times-more-powerful-than-h20-in-fp4
Huawei Atlas 350 AI accelerator on Ascend 950PR delivers 1.56 PFLOPS FP4, surpasses Nvidia H20. Comparison of specs, HBM memory, price. For AI developers — details and nuances.
ROCm is catching up to CUDA in AI: OneROCm and Triton https://sudonull.com/rocm-is-catching-up-to-cuda-in-ai-onerocm-and-triton
AMD unified ROCm into OneROCm for all accelerators. Triton writes kernels for AMD/Nvidia. vLLM inference on par with CUDA. Six-week releases. Test the stack on Strix Halo.
Cerebras IPO: profit, OpenAI and wafer-scale chips https://sudonull.com/cerebras-ipo-profit-openai-and-wafer-scale-chips
Cerebras turned profitable and is preparing for IPO with an OpenAI contract for $20 billion. Analysis of architecture, risks, and prospects for the AI accelerators market.
Apple M5 chip: 3nm process, 128 neural engine cores, AI revolution https://sudonull.com/apple-m5-chip-3nm-process-128-neural-engine-cores-ai-revolution
Apple M5 with 3nm TSMC and 128 neural cores: AI 4x faster than M4. Why the chip is not needed by 95% of users, but Apple sells it to everyone. Detailed analysis.