ComfyICU / infrastructure evidence
Local GPU infrastructure
Only fixed-protocol runs: model-file pages evicted before every cold process, process block reads verified, then a different-seed warm run in the same process. Schema v4 captures block-device snapshots; cgroup visibility determines whether their deltas are attributable to the benchmark.
Dynamic: DuckDB serves the materialized normalized run table, so the UI never needs to be regenerated.
Checking status…
Storage path vs cold penalty
model-file GiB/s → cold − warm seconds| Run | Storage | Cold | Warm | Penalty | Model read | Comfy read | fio QD1 / QD32 | PCIe H2D | Evidence |
|---|---|---|---|---|---|---|---|---|---|
| Loading DuckDB results… | |||||||||
Measured economics / all GPU classes
Rental and purchase Pareto
Every performance point comes from a successful five-pair schema-v3-or-newer run. Pareto uses same-process warm throughput to remove cold disk/model loading. “Resident” does not guarantee that the whole model fits in VRAM: when ComfyUI offloads through host RAM/PCIe, that cost remains in the measured warm result. Rental uses the Vast price captured with that exact host. Purchase prices are explicit, editable inputs joined to the same measured throughput.
Normalize the quote scope before buying. Some inputs are card-only marketplace asks while “White Box” supplier quotes may include more of a system. Capex rankings are provisional until every quote uses the same bill of materials.
Rental price vs measured throughput
lower cost → / higher throughput ↑| Loading economics… |