BENCHMARK INTELLIGENCE
Benchmark Results
Explore published AI system results with their model, workload, system boundary, software stack and source evidence intact. This is a source-linked result registry, not a universal leaderboard.
FILTER THE EVIDENCE
Find comparable submissions
Advanced workload filters · InferenceX
PUBLISHED RECORDS
— results
Loading result registry…
| Submitter | Accelerator / system | Model / scenario | Result / selected metric | Status / evidence |
|---|
Try a broader filter or reset the filters.
DATA SOURCES
What is in the registry?
COMPARABILITY POLICY
Context before ranking.
Do not compare results across different benchmark suites, models, scenarios, divisions, system types, units or accuracy targets. For MLPerf, preserve the suite version, Closed/Open division, Available/Preview/RDI status, scenario, inference mode, accuracy target and full system boundary. Inferred submissions are marked separately.
The MLPerf v6.1 summary snapshot contains no measured power records. Design power is not measured power. For InferenceX, median_intvty is median interactivity in tok/s/user (higher is better), not latency; p99_itl and p99_tpot are latency metrics in seconds. Throughput metrics are tokens/s per physical GPU unless the source key states otherwise.
When a normalized metric is selected, the table filters to records that report that metric and displays its common unit. Metric sorting respects the source-defined optimization direction (higher or lower is better); context-only metrics are not ranked. This is not a cross-suite score. Compare InferenceX rows only after matching model, workload type, ISL/OSL, concurrency, precision, framework/image, speculative mode, aggregated vs. disaggregated topology and prefill/decode parallelism. GPU power/energy values require power_valid=1 and power_metric_schema_version=2. CPU/module telemetry has an independent cpu_power_valid verdict. An absent CPU verdict means there is no CPU-side measurement leg, not a failed GPU measurement. The importer preserves failed, incomplete and unreported states separately; do not combine component boundaries without checking their definitions.
Every row links to its publisher's source record where available. A public submission or observation is attributed to its publisher; it is not automatically an independent replication.