AMD跑DeepSeek性能超H200！128并发Token间延迟不超50ms，吞吐量达H200五倍
在稍早一些的测试当中，使用SGLang在MI300X上运行DeepSeek-R1，仅过了两周就相比于day 0时性能提升到了4倍，吞吐量达到了每秒5921 Tokens。 [1] https://rocm.blogs.amd.com/artificial-intelligence/DeepSeekR1-Part2/README.html ...
来自MSN1 个月
AMD Analyst Is ‘Nervous’ About the Chip Maker. Here’s Why.
Chip maker Advanced Micro Devices might soon suffer from bloated chip inventory levels at retailers and distributors, Bernstein warns. Industry shipments of central processing units, or CPUs, were ...