ComfyUI user weighs zram weight offloading: zstd decompression may bottleneck the inference hot path
Agreeable_Return2632 · reddit · 2026-09-27
A user running MiniMax H3 in ComfyUI with CPU offload (i5-14400 + RTX 4070 + 32GB RAM) considers zram to extend usable memory for offloaded layers. Rough math suggests zstd gives only 1.3-1.4x compression on near-random-entropy fp16 weights, adding 10-15GB effective capacity.
The catch, which the author flags themselves: offloaded layers are read on every forward pass, so decompression runs continuously in the hot path — RAM bandwidth is 35-50GB/s while zstd decompression tops out around 5-10GB/s even parallelized. They ask whether anyone has actually benchmarked zram for weight offloading in inference pipelines.
More from Infra
- Running Qwen2.5-Coder-3B on Android: Jupyter + Pydantic Evals on-device eval pipeline — alex_b3b · 2026-11-11
- 22% of Nvidia revenue now books to Singapore, but only ~1% is local buyers — Gregory_C_Allen · 2026-11-11
- Huawei's Kirin 9050 NPU claims offline 30B MoE on smartphones, +140% perf — AIFlow_ML · 2026-11-11
- Stratechery: Intel vs AMD's diverging earnings and Intel's need for an old-new path — Stratechery · 2026-11-11
- a16z Podcast: Mike Volpi on What the 1990s Internet Boom Teaches Us About the AI Buildout — a16z Podcast · 2026-11-11
- Alibaba files to list chip-design unit T-Head in Hong Kong to fund AI push — pstAsiatech · 2026-11-11