AMD MI355X Hits 1.7x Perf-Per-Dollar vs DGX B300 via SGLang KV Cache Fix
zephyr_z9 · x · 2026-09-23
SemiAnalysis says AMD's MI355X delivers up to 1.7x better performance per dollar than DGX B300.
AMD's Mainland China UMBP team, in collaboration with Alibaba and the SGLang community, added a feature to SGLang that removes duplicated KVCache between local L2 DRAM and distributed L3 DRAM, allowing up to 2x more KVCache to be stored in DRAM. The feature is called UnifiedRadixCache external cache.
SemiAnalysis stresses this marks a trend of AMD increasingly being first-class co-designed for new features in widely used production engines like SGLang.
More from Infra
- Intelligence may become abundant before power plants and fabs can keep up, argues thread — r0ck3t23 · 2026-09-23
- Cloudflare CTO Dane Knecht makes TIME's exec list, rebuilding web for AI agents — michellechen · 2026-09-23
- Parakeet.java ports NVIDIA's Parakeet ASR to pure Java, transcribing 15-min speech in ~15s — mukel90 · 2026-09-23
- Nvidia-backed Nscale omits its biggest customer Bytedance from US IPO filing — The Decoder · 2026-09-23
- New Qwen3.8 27B quants beat ISTA and Unsloth byte-for-byte on three corpora — Dutchnamn · 2026-09-23
- crabbox now runs on boxd: isolated KVM microVMs with ms boot times for repo commands — steipete · 2026-09-23