897M tokens and 30.5TB of VRAM swapping in one month, all running locally
Distinct-Pie2389 · reddit · 2026-11-14
A Reddit user shared 30-day stats from a fully local setup: 897M tokens processed (888.27M prompt + 8.75M output) and 30TB of VRAM safely swapped across llama-swap, strata, ninfer, ComfyUI, and training jobs.
He notes cached prompt tokens are included in the processed volume — they represent reused input, not necessarily freshly computed prompt processing.
More from Infra
- Dario Amodei: compute planning is Anthropic's 'deepest business dilemma' amid revenue unpredictability — shiringhaffary · 2026-11-14
- From-scratch Rust+CUDA inference engine for DeepSeek hits 1740 TPS on 4x B300 — saranormous · 2026-11-14
- The Bank of Nvidia: ZeroHedge maps $100BN+ in circular AI deals after Huang exit — SumitGup · 2026-11-14
- Alleged $7.5B Nvidia chip pipeline: a 10% cut would mean nearly $1B — DesireeCachette · 2026-11-14
- Still waiting for a convincing, openly disclosed pretrain run on Ascend chips — teortaxesTex · 2026-11-14
- AI Now Institute: BAAJA is a 'backdoor bailout' fast-tracking AI data center buildout — AINowInstitute · 2026-11-14