TokenSpeed hits SOTA for Kimi K3 on AMD MI355X: 1.4x faster than ATOM at 216 tok/s
zhyncs42 · x · 2026-10-10
TokenSpeed claims new SOTA serving Kimi K3 on 8x AMD Instinct MI355X: 216 output tok/s per user, 1.4x the decode speed of ATOM at C1 while running higher precision, measured on a real agentic workload (multi-turn SWE-smith, 50K+ context, 10–15 turns). All AMD kernels were rewritten in Triton + Gluon with the AMD Triton team; a technical deep dive is promised.
More from Infra
- Full AI video made at home on a single RTX 3090 with Qwen, FLUX, MiniMax and YuE2 — solyarisoftware · 2026-10-10
- Baseten launches Project Beacon with Goodfire for inline open-model safety — baseten · 2026-10-10
- Huawei launches OceanStor M900 context memory storage for hyperscale AI data centers — pstAsiatech · 2026-10-10
- Morgan Stanley: Amazon Trainium3/4 to drive Intel EMIB-M packaging, capacity scaling 8x by 2028 — pstAsiatech · 2026-10-10
- Data centers could become SpaceX's biggest revenue stream, with $1T revenue on the horizon — Dr_Singularity · 2026-10-10
- New gTLD application for .lan shows why internal services need real registered domains — evilsocket · 2026-10-10