Homelabber uses second-gen Optane memory sticks to run a 2.8-trillion-parameter AI model on CPU alone
Reported by Tom's Hardware1 source ↓
A Reddit homelab user ran Moonshot AI’s 2.8-trillion-parameter Kimi K3 CPU-only on Optane PMem 200 at 0.85 tokens/s with a 1-bit quant.
Key points
- The build uses a 40-core third-generation Xeon Scalable CPU and second-generation Optane PMem 200
- The model ran at 0.85 tokens per second with 1-bit quantization; 3-bit quantization ran at 0.75
- The builder used dual 12W blower fans to cool the Optane modules
- The builder plans to add GPUs and use DRAM as a faster memory tier
Details
Tom’s Hardware reports that the builder used Optane in App Direct mode, which exposes it as a separate persistent region that software can map directly. The builder said the modules became extremely hot during sustained inference and warned that active DIMM cooling is necessary.
The reported performance figures are pre-tuning, without adjustments to thread counts or NUMA affinity. For comparison, an earlier Optane build ran a smaller, 1-trillion-parameter model at about 4 tokens per second, but used an RTX 3060 alongside the CPU.
Intel specifies forced airflow for PMem 200 modules. The builder plans further benchmarks with GPUs and some model data loaded into DRAM.
Written from coverage as of
Images
- Tom's Hardware
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh