➀ SambaNova introduces SN40L, an AI chip system supporting 5 trillion parameter models, with performance 3.1x and 2x that of NVIDIA's H100 in inference and training, respectively, at only 1/10 the cost. ➁ The SN40L chip, built on TSMC's 5nm process, features 1020 billion transistors and 1040 custom 'Cerulean' RDU cores, offering 638TFLOPS (BF16). ➂ SambaNova's system includes 520MB of on-chip SRAM, 64GB of HBM, and 1.5TB of external memory, enabling support for large-scale model training and inference.
Related Articles
- AI bubble conundrum29 days ago
 - Make Alibaba Great Againabout 1 month ago
 - China objects to US chip hand-me-downsabout 2 months ago
 - China bans its biggest tech companies from acquiring Nvidia chips, says report — Beijing claims its homegrown AI processors now match H20 and RTX Pro 6000Dabout 2 months ago
 - Nvidia tipped to be TSMC's first 16A customer, ahead of Apple — Feynman GPUs could make full use of GAA transistors and backside powerabout 2 months ago
 - Nvidia Rubin CPX die shot reveals graphics-specific hardware blocks not needed for an AI GPU — Rubin CPX may form the foundation of next-gen RTX 6090about 2 months ago
 - The new frontier: Meet the power players cashing in on the AI gold rushabout 2 months ago
 - Nvidia claims software and hardware upgrades allow Blackwell Ultra GB300 to dominate MLPerf benchmarks — touts 45% DeepSeek R-1 inference throughput increase over GB200about 2 months ago
 - ‘Still, you’re paying for dinner’ Nvidia CEO shoots back, after TSMC CEO jokes about his billionaire status2 months ago
 - White House peddling export licences3 months ago