Scaling Kimi K3 on H200s: Engineering Insights from 1000+ Chips
hsu_byron · x · 2026-08-01
Author shares practical tuning insights for high-throughput inference of Kimi K3 on H200 clusters:
- TP+EP beats DP+EP: With only 1/4 of layers using gated MLA, DP attention provides less benefit and increases latency.
- FP8 KV cache works well: No quality or throughput regression measured, effectively doubling cache capacity.
- TP32+EP32 boosts capacity: Provides 4M KV-cache tokens per replica vs 800K with TP16+EP16.
- Disable radix cache for low hit-rate workloads: Saves Mamba-state slots.
- Non-speculative decoding wins at high concurrency: Overhead outweighs savings above 64 concurrent requests.
The setup has run reliably at full load across 1000 chips for days, though some multi-node instability issues are still under investigation.
Related event: Kimi K3 Throughput Optimization on H200 Clusters(3 posts)→
More from Infra
- NVIDIA quietly delists the entry-level RTX 5050 from its marketplace, following the RTX 5090 — chemist_slime · 2026-10-16
- Grok claims 816B MoE model runs locally on dual M5 Ultra Mac Studios via MLX — grok · 2026-10-16
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16