Kimi K3 Inference Math: 77k-Context MLA Costs ~386B FLOPs-Equivalent, GPUs Not Enough
YouJiacheng · x · 2026-10-26
YouJiacheng argues Kimi K3 inference is not memory bound and that even at full MFU, Kimi's GPUs are insufficient — the "6x" figure is only a lower-bound estimate.
- At 77k context, K3's MLA costs roughly 77k×24×96×(512+576)×2 = 386B activation-parameter-equivalent compute (the ×2 because MoE runs MXFP8 while MLA runs BF16).
- MoE and linear projections hit compute bound (core attention's bound status is batch-size independent); FP8 KV's real value is avoiding OOM at large batch sizes.
- FP8 KV dequants to BF16 compute (per FlashMLA); 96-head MLA has an arithmetic intensity of 2×96×(512+576)/576 ≈ 363 FLOPs/Byte — compute bound even without speculative decoding.
More from Infra
- Sail equivalence proof and riscv-coq cross-checks uncover real RISC-V spec bugs — xennygrimmato_ · 2026-10-27
- xAI's Colossus 1 hits ~14.6 exaflops FP64, ~5x Top500 leader — cis_female · 2026-10-27
- Lambda runs 19 B200 nodes in a 16-node power budget, boosting throughput 24% — TheZachMueller · 2026-10-27
- Podcast: Micron's HBM roadmap, Meta's AI infra, and NVIDIA's supply chain lock-in — BenBajarin · 2026-10-27
- SGLang MegaKernel + PD Pushes the TCO-Latency Frontier on GB300, SemiAnalysis Benchmarks Show — AccBalanced · 2026-10-27
- Scaling self-hosted Convex: compiling TypeScript Convex functions to WASM — leventov · 2026-10-27