Kimi K3 Training Optimizations Save ~90GB HBM Per GPU
AccBalanced · x · 2026-08-30
At Kimi K3 scale, memory directly dictates the number of GPUs required for training. Key optimizations include:
- Streaming Gradients: Used for Adam updates, reducing peak host memory by 33%.
- SiTU-GLU Activation Optimization: Native autograd saves redundant tensors, consuming >100GB HBM at 100k tokens/GPU. A custom streaming operator using a fixed-size workspace saves 90GB of HBM per GPU.
Related event: Kimi K3 Full Fine-Tuning Launches on AC2 with 40% Fewer GPUs(3 posts)→
More from Infra
- NVIDIA launches RTX Spark for local AI on Windows PCs — ostrisai · 2026-10-16
- SkyRL ships Day-0 RL support for Beam with +62% LoRA decode throughput in vLLM — vllm_project · 2026-10-16
- DFlash2 delivers up to 3.03x decode-only speedup for Beam across five benchmarks — vllm_project · 2026-10-16
- vLLM's Beam serving kernels cut GB200 forward time from 8.15 to 6.47 ms at TP8 batch 128 — vllm_project · 2026-10-16
- Reflection's 501B MoE Beam hits 463 tok/s on 8x GB200 with NVFP4 at day 0 — zhyncs42 · 2026-10-16
- Gemma 4 31B QAT on one 24GB card: 46 tok/s plain, ~80 with draft model on RTX 4090 — Subject_Poet_3861 · 2026-10-16