Kimi K3 on AMD MI355X: +38% agentic throughput, +86% peak vs NVIDIA B300
YouJiacheng · x · 2026-10-27
Lorenz Research reports end-to-end optimization of Kimi K3 agentic inference on a single 8×AMD MI355X node (kernels, KV cache, communication, speculative decoding), benchmarked with SemiAnalysis InferenceX AgentX harness replaying real coding-agent sessions:
- Without speculative decoding: +38% throughput vs NVIDIA B300 (17,389 vs 12,566 tok/s/GPU at c80)
- With speculative decoding (DSpark + ReplaySSM): 23,330 tok/s/GPU, +86% over B300
- At GPU-hour prices ($1.50 MI355X vs $2.26 B300), that's 56.0M tokens per $1 — 2.80× B300
- Model quality essentially unchanged; passed the Kimi K3 Verifier test
Key claims: for specific models and workloads the bottleneck is software, not chips; model-specific end-to-end optimization yields large gains; AI-assisted engineering is helping non-CUDA platforms close the gap with CUDA.
More from Infra
- YC-backed Isoquant claims 1.2T tokens and 1,000+ customers in its first month — ycombinator · 2026-10-27
- Off-grid data center gas power emits 63-80% more than the grid — 2C_ornot2C · 2026-10-27
- Runware puts FFmpeg in the cloud: common commands via API from $0.02/GB-second — aziz4ai · 2026-10-27
- Google hiring anti-scraper analyst at up to $216K/year to guard search results from LLMs — gaganghotra_ · 2026-10-27
- miniTE Brings Stable Native NVFP4 RL, 1.39x Faster Than BF16 on B300 GPUs — InfiX-ai · 2026-10-27
- SemiAnalysis: Anthropic S-1 shows inference is great business, neoclouds still thrive — johncoogan · 2026-10-27