Explore AI inference performance using MLCommons’ official results, and consult the Sistema Vectorial (SV) hardware tests separately. Meaningful comparisons depend on the workload, system configuration and measurement conditions.
External reference · MLCommons
MLPerf® Inference: Datacenter
The official interactive results viewer compares complete systems running defined inference workloads. Its results describe the tested hardware and software together; they do not establish a universal ranking of individual CPUs, GPUs, TPUs or NPUs.
Start with a benchmark round, model and scenario. Then inspect the division, availability, accelerator count and system details before comparing results. Use the viewer’s filters and full-screen control; on narrow screens, scroll horizontally.
Updates on opening or reloading. This page loads the official MLCommons dashboard from Tableau Public. The benchmark round and source publication date appear inside the viewer. A fresh visit may show the same edition until MLCommons publishes or revises its results.
The viewer is provided by MLCommons through Tableau Public and requires access to that external service. If it does not load, use either link above. No locally stored reference table is presented as a live update.
How to read and compare these results
- Keep the test comparable. Compare the same benchmark and scenario from compatible versions. Closed division results use the reference model; Open division permits changes that must be considered separately. Available, Preview and Research/Development/Internal systems describe different availability conditions.
- Read the complete configuration. Check processors, accelerator model and count, number of nodes, software and numerical precision. A result from a large installation is not the performance of a single accelerator.
- Keep the measurements distinct. Throughput measures completed work per unit of time; latency measures response time. Compare matching units and quality requirements. Energy or efficiency comparisons require the corresponding measured system power; a nominal power rating is not a substitute.
The viewer retains the source’s result identifiers and submission details. These external measurements are not SV tests and do not confer SV approval or a recommendation for a particular deployment.
Source, revisions and attribution
Data and visualisation: MLCommons. The embedded dashboard is supplied through Tableau Public. Benchmark version, division, scenario, submitting organisation, tested system and result identifiers should be read together with each result.
MLCommons may correct or invalidate published results. Consult the official results change log, the benchmark rules and the results usage guidelines when citing or comparing an entry.
MLCommons and MLPerf are trademarks of MLCommons Association. This page provides an independent editorial introduction and access to the official viewer. No affiliation with or endorsement by MLCommons is implied. The external data and visualisation retain their own applicable terms independently of this website’s general licence.
Independent evaluation · Sistema Vectorial
SV hardware tests and evidence
The SV hardware page distinguishes manufacturers’ specifications from documented physical tests, and explains the workload, observed results and limits of each case. A shared ranking requires a common evaluation campaign and comparable conditions.
Consult the current status and evidence on the SV page. MLPerf results are not combined with SV measurements to produce a joint score or ranking.
The SV page is maintained separately and may be in Spanish.