Perplexity (@perplexity_ai)

GB200과 H200의 성능 비교 결과를 제시했다. NVLS all-reduce 지연시간과 MoE prefill/ combine 시간이 크게 줄었고, 디코드 단계에서도 고토큰 속도에서 더 높은 처리량을 유지해 대규모 모델 추론 성능 개선이 확인됐다.

https://x.com/perplexity_ai/status/2054204425833726353

#gb200 #h200 #benchmark #moe #inference

The introduction of the Vera Rubin platform shifts the calculus for AI infrastructure planning. As the industry moves toward HBM4, understanding hardware refresh cycles becomes a core component of fleet optimization.

While H100 and Blackwell GPUs remain key workhorses, secondary-market demand for current-gen accelerators has reached a unique inflection point. This analysis explores the technical and financial variables influencing hardware transitions as the industry prepares for the Rubin wave.

https://www.buysellram.com/blog/the-post-gtc-gpu-market-shift-when-to-liquidate-h100-h200-and-blackwell-assets/

#NVIDIA #TechStrategy #DataCenter #GPU #GraphicsCard #GPULiquidation #H100 #H200