Rented GPU Bills: Host CPU and Script Defaults Made Costs 31x Higher
Worldly_North_7213 · reddit · 2026-09-11
From 15 self-funded rental jobs ($33.60): an SDXL LoRA job ran 1.85x slower on a 5-vCPU host than 24 vCPUs (same 4090), and an H100 with 16 vCPUs was slower than the desktop 4090 host at 4x the rate — fixed with dataloadernumworkers. Meanwhile fp32 batch-1 defaults on H100 cost $0.0112/image vs $0.00036 with fp16 batch-4 on a 4090 spot instance, a 31x gap, with GPU showing 99% utilization throughout. Lesson: the meter runs at card hourly rate regardless of useful work.
More from Infra
- SemiAnalysis: OpenAI reportedly negotiating ~1GW TPU commitment while seeking sites — dylan522p · 2026-10-28
- I tested 11 video upscaling methods for AI video: local, Topaz and cloud compared — lopas99 · 2026-10-28
- Thermo-backed CTO says AI is in its dial-up phase, touts disaggregated compute for an ultrafast 'broadband moment' — beffjezos · 2026-10-28
- DeepSeek-V4.1-Flash Q2 runs at 41 tok/s on one RTX 3090, 90% token match vs API — musicurgy · 2026-10-28
- USearch v3 RC ships after 1,000 days, its libraries now embedded in 550K-star projects — viglovikov · 2026-10-28
- SK Hynix says memory demand will outpace supply through 2030 — Chemical_Hawk_6307 · 2026-10-28