Speeding Up Inference via SSD Streaming of Experts
gajesh · x · 2026-07-19
Someone proposed a cost-effective approach to run Kimi K3 at 1t/s: when VRAM is near its limit, you can use SSD streaming of experts to load certain experts on-demand from an SSD.
They also noted that Mac machines boast SSD read speeds up to 7GB/s, implying that this setup is practically viable in certain scenarios. The attached image shows Micron's data center SSD page, highlighting the importance of high-bandwidth storage for AI inference and training.
More from Infra
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27
- NVIDIA claims HGX Rubin NVL8 delivers 10x tokens per megawatt vs B200 — thursdai_pod · 2026-10-27
- NVIDIA Claims HGX Rubin NVL8 Delivers 10x More Tokens per Megawatt Than B200 — thursdai_pod · 2026-10-27
- Together lets multiple LoRA post-training experiments share the same GPU capacity — togethercompute · 2026-10-27
- Commoditization dilemma: either human labor or GPUs become the commodity — _AustinCalvert_ · 2026-10-27