Specification and Vast score
—Useful before rental, but it does not observe the mounted path used by ComfyUI.
ComfyICU / RTX 5090 cold-start analysis
Storage path & cold start
Why nominally similar RTX 5090 hosts load models differently—and how to separate pre-run screening evidence from diagnostics measured inside the workload.
Current interpretation
Validate the complete path, not the SSD label
Cold-start variance follows the path that actually delivers model bytes to ComfyUI: drive, negotiated link, filesystem/container layers, host state and loader access pattern.
The owned nodes use one 2 TB 9100 Pro model drive plus a separate OS drive, without model-cache RAID. The repeated application result and physical-read evidence—not a synthetic threshold—accept each node.
—buffered fio vs MiniMax cold penalty—the strongest current independent screening relationshipLoading the in-workload diagnostic and dataset timestamp…
Loading DuckDB analysis…
Evidence hierarchy
Read the measurements in this order
Each layer answers a different question. Stronger proximity to the timed run also means less independence from the outcome.
Buffered fio after invalidation
—Best current independent proxy because it traverses more of the real buffered path.
Active workload read rate
—Explains where elapsed time went, but is measured inside the timed cold process and cannot screen an untested host.
Admission rule: only the complete repeated ComfyUI workload accepts a node. Independent probes help find a bad path; in-run diagnostics help explain a failed result.
The central multivariable result
The broad storage signal weakens inside healthy link classes
Spearman correlation with MiniMax cold−warm penalty. Negative means faster storage tends to accompany a smaller penalty.
| Measurement | All single paths | Exposed NVMe | PCIe 5.0 x4 | PCIe 4.0 x4 |
|---|
Apparent contradictions explained
Why a slower probe can accompany a smaller penalty
The standalone reader and ComfyUI do not always realize the same delivery rate. One additional host is a noisy accelerator/runtime outlier.
Keep raw time visible: machine 45827 has a small incremental penalty but extremely slow cold and warm wall times. A small `cold − warm` value does not necessarily mean fast service.
Historical RAID decision evidence
RAID was not required for the fleet design
Loading the historical purchase-relevant MiniMax slice…
| Historical slice | Hosts | Machines | Standalone reader | Penalty mean | Warm mean | Cold mean |
|---|
Automatic statistics
Screening evidence and in-run diagnostics
Metrics are grouped by when they become available. Pearson measures linear association; Spearman measures rank ordering.
| Metric | Role | N | Pearson r | Spearman ρ |
|---|
Multivariable single-path explorer
Why are some “single SSD” hosts slow?
Slice the latest non-RAID MiniMax hosts across the SSD link, queue/request limit, root-port depth, buffered loop, controller, GPU transfer path, kernel and runtime. Each filter compounds with the others; no cross-host group is presented as a controlled intervention.
More filters 10 available
| Selected group | Hosts | Machines | Standalone reader mean / range | Active read / duration | Penalty mean / range | Buffered fio | H2D |
|---|
| Host | Drive / SSD link | CPU attachment | Loop / runtime | GPU path | Buffered fio → standalone reader | In-workload active read | Penalty / active-duration residual | Leading observed limitation |
|---|
Interpretation: one upstream PCI hop is a root-port-like path and is labelled likely CPU-direct. It is not proof of motherboard routing. “No active buffered loop observed” requires cgroup-attributed block layers; older or host-wide evidence is shown as unresolved. Active-read rate and duration are measured inside the timed cold workload. The displayed residual is `active duration − cold penalty`; it is not an independently observed overlap interval.
Disk × setup × performance
Compare the exact installed path
Every row is the latest fixed-protocol result for one physical machine and workload. Disk identity stays beside negotiated storage topology, CPU/GPU path and application performance so a product name is never interpreted without its installed setup.
| Disk name | Storage setup | CPU / GPU setup | Selected performance | Cold / warm / penalty | Standalone reader / buffered fio | Machine / result |
|---|
Reading this view: the advertised disk identifies the workload volume where Vast exposes it; “visible devices” preserves every NVMe name reported by the host. One upstream PCI hop is only a likely CPU-direct path, while multiple hops suggest a chipset or switch. The table does not pretend those labels are controlled interventions.
Repeated cross-workload evidence
Two workloads on comparable 575 W hosts
The paired cohort is limited to single-drive and RAID 0 layouts. This remains a cross-host comparison, not a RAID-only intervention.
| Host | Layout | GPU path | MiniMax cold / warm / penalty | Krea cold / warm / penalty | Model reader |
|---|
Broader admitted slice
MiniMax application penalty vs model-reader rate
575–600 W, H2D ≥10 GiB/s, full-width x16 RTX 5090 hosts. Each point is one physical machine.
Hover, focus, or tap a point to identify its physical machine.
| Layout | Hosts | Model reader mean | Penalty mean | Penalty median | Penalty range |
|---|
Robustness check
Leave one physical host out
Positive means RAID appears faster; negative means the remaining RAID observation is slower.
Reproducibility
Cross-workload and reacquisition controls
Strong agreement across Krea and MiniMax distinguishes persistent host-path behavior from five-pair measurement noise.
| Machine | Workload | Previous schema | Current schema | Previous penalty | Current penalty | Change | Reader previous → current |
|---|
Evidence hierarchy
Physical block-path controls
Cgroup-attributed device deltas are stronger than host-wide counters. Neither converts a cross-host comparison into a causal SSD experiment.
| Host | Layout / media | Evidence class | GPU / H2D | Model reader | Cold / warm / penalty | Raw result |
|---|
Protocol: ComfyUI v0.37.0; five verified cold→warm pairs; physical machine is the independent unit. Reacquisitions demonstrate repeatability but are deduplicated in cross-host summaries. See the standalone decision document for interpretation and the full experiment journal for provenance.