alt.hn

5/30/2026 at 8:17:49 PM

768GB Intel Optane DIMMs to run 1T-parameter LLM with single GPU at 4tps

https://www.tomshardware.com/tech-industry/artificial-intelligence/enthusiast-runs-1-trillion-parameter-llm-from-768gb-of-intel-optane-dimm-memory-sticks-local-kimi-k2-5-install-achieved-roughly-4-tokens-per-second

by walterbell

5/31/2026 at 2:28:32 AM

Ah Optane, what might have been...

Even over PCIe, I imagine the advantage vs. NVMe is lower latency and more operations per second.

by musicale

5/31/2026 at 1:22:27 AM

The bottleneck in this setup is PCIe bus. You don't need optane to saturate it. 4 regular SSDs might do just fine.

by lostmsu