对NVDA有兴趣的同学,建议听听老黄在2023 NVIDIA GTC keynote, 信息量很大

根据老黄讲的, Moore's law is about to die because the transistor size in CPU is approaching its physical limit (CPU accounts for 80% of current data center processors), therefore it is not practical and economical to increase the scales of CPU to meet AI demand as the energy consumption by data centers is limited. The only solution for this is to accelerate computing power while lowing energy consumption. NVIDA's A100 and H100 chips are designed by linking CPU (based on ARM architecture) with GPU together to optimize computational power while using only a fraction of energy by CPU serves. The next generation H200 was already announced last year and will be in production in 2024. The 2024 GTC will be on March 18.

所有跟帖： 

• The computational speed of NVIDA GPU increased 1 million -Outlook2024beyond- ♂ (36 bytes) () 02/26/2024 postreply 14:59:42

• 这个keynote油管上有 -Outlook2024beyond- ♂ (0 bytes) () 02/26/2024 postreply 15:17:25