Hold the workload accountable.
Compare the same work and quality requirements. Show latency misses, failures and the resource denominator alongside throughput.
GPU UTILIZATION / OPEN EVIDENCE
More utilization.
More work per GPU.
Put every GPU-hour to work. Explore the software, scheduling and runtime choices that turn available capacity into useful output.
Explore the evidence01 / COMPARE
The utilization explorer
Historical measurements · IdleBench v0.1
Loading published evidence…
HARDWARE / LATENCY TARGETNVIDIA A10 · 15 s
USEFUL OUTPUT × TAIL LATENCY
↖ More work, lower latencyCapacity has a tradeoff.
Nodus configurationsBaseline configurationsDedicated references
Configuration results
| Configuration | Requests / GPU-h Within 15 s · higher is better | p95 latency Seconds · lower is better | Within 15 s | Within 5 s | Evidence |
|---|
No matching configurations
Try another search or reset the filters above.
Historical descriptive results, not a universal ranking. One traffic seed per row; baseline and imported Nodus measurements used different A10 hosts. Stock dedicated references extrapolate short per-model runs; imported reference allocation details remain incomplete. Read the scope.
RESEARCH COLLECTION
Published studies have their own workloads, baselines and limitations. They are not combined into a score or treated as newly reproduced results.
02 / WHAT WE MEASUREBusy is a signal.
Busy is a signal.
Useful is the outcome.
GPU activity explains what the hardware is doing. Useful work per GPU-hour tells you what that activity delivered.
Separate the kinds of capacity.
Memory freed, device time available, useful work completed and provider time billed answer different questions.
Make every result inspectable.
Preserve the workload, baseline, source revision and limits. Report wins and losses. A research result does not mean a feature is shipped or enabled.
IdleBench v0.1: scope, limitations and reproduction status +
The serving explorer is a direct projection of the published suite s2 table. Each row uses one traffic seed; it cannot establish between-run confidence. Bootstrap intervals in the paired source table are within-trace comparisons, not independent seed replication.
Stock baselines use vLLM 0.30.0. Model IDs and revision pins are not present in the source table; exact model parity, including a reported Qwen2.5-3B revision mismatch, remains a limitation. The imported Nodus agent cold-start baseline is labelled separately. Stock vLLM sleep is a strong baseline and wins many latency comparisons. Snapshot restore answers a different resource-release question; this table alone does not prove billed savings or successful use of the released GPU by another workload. Host-to-host differences and caching matter.
The new stock dedicated-per-model reference extrapolates each model’s first 60 seconds. It is not a simultaneous full-trace run on K GPUs. It is excluded by default and labelled when shown.
The raw-event archive and full GPU execution kit are not yet packaged for public download. Existing v0.1 results remain historical evidence; the v0.2 protocol and scorer are drafts, not new measurements. No provider leaderboard or generic “Nodus wins” score is inferred.
03 / TAKE THE EVIDENCE
Open the numbers.
Source tables, methodology and file hashes.
No account required.
MD / MEASURED
IdleBench source tables
Original serving results and paired comparisons.
↗JSON / DERIVED VIEWExplorer dataset
Table rows, source-line pointers and research index.
↗v0.2 · DraftBenchmark protocol
Workload, allocation and quality requirements.
↗JSON / PROVENANCEDownload manifest
Source revisions and SHA-256 for every download.
↗FROM RESEARCH TO YOUR POOL
Open Nodus console ↗Make the next GPU-hour count.
Explore your pool’s efficiency and available setup paths in the Nodus console.