SoAI Bench · leaderboard
GPU leaderboard
Certified SoAIBench results published to soai.to, grouped by
GPU model. Each model is ranked on the median of its published runs, not on its best single
run. Open a model to read every run behind that median, with the lowest and highest score
those runs produced.
| # | GPU model | Vendor | Runs | Median overall | Compute | Memory | Latency |
|---|---|---|---|---|---|---|---|
- Median overall score
- Lowest published run
- Highest published run
- Highest minus lowest
Score ranges · each track runs from zero to the highest median on the board
| Published | Submitted by | Overall | Compute | Memory | Latency | Variance | Driver | Platform | CPU | RAM | SoAI | Max temp | Avg power |
|---|
SoAI Bench · scoring
What a score means
SoAIBench runs synthetic OpenCL workloads on one selected GPU and reduces
them to four numbers. Only a certified run on the standard profile can be ranked, and
soai.to recalculates every number from the submitted evidence before it appears
here.
-
Warm-up, then five measured passes
A certified run completes at least 120 active seconds of warm-up and then five measured passes. Its published score applies the scoring formula to the median compute throughput and median memory bandwidth from those passes. Variance, drift and available temperature and power readings remain descriptive diagnostics; they do not impose an arbitrary agreement or sensor-availability threshold.
-
How the overall score is built
Overall score is
compute × 0.40 + memory × 0.60. Compute measures throughput in GOPS and memory measures bandwidth in GB/s. Latency reports kernel-dispatch performance as a diagnostic score and does not contribute to the overall score. Higher is better in all four columns. -
What validation proves
soai.torecomputes each pass, the medians, the variance, and the eligibility rules from the uploaded evidence, and rejects a submission whose own totals disagree with that recalculation. A validated result is therefore internally consistent with the current SoAIBench scoring protocol. Nothing in it proves which binary ran on which card.
SoAI Bench · publishing
How a result gets published
Share a completed certified run from SoAI after reviewing its exact public payload. The result dialog provides its public URL, an open link, and a copy action.
-
Certified runs only
In SoAI you open Hardware, pick one GPU, and start a certified SoAIBench run on the standard profile. Quick runs and stress runs stay in your local history and are never ranked.
-
No account
There is nothing to sign up for. Before anything leaves the machine you see the exact payload: GPU model, vendor, driver, OpenCL runtime, the five passes, telemetry, and a display name. Sharing from SoAI uses Anonymous. A public result never carries the GPU UUID, the PCI address, the hostname, the IP address, or any local device identifier.
-
Permanent public submissions
Published scores cannot be withdrawn by contributors. You can delete terminal local history in SoAI; this never changes the public result. Webmaster moderation can exclude, restore, or remove invalid or problematic results.
-
Every genuine run counts
Repeated runs from the same hardware, including equal scores, are independent samples. Only retransmitting the same publication is deduplicated. The distribution includes every accepted, publicly ranked run. Ranking first takes the median of each installation's runs, then gives every installation one contribution to the model median.

