SoAI Bench · leaderboard

GPU leaderboard

  • Top GPU —
  • — GPU models
  • — Runs
  • — Installations

Published GPU models, ranked by median score
# GPU model Vendor Runs Score Compute Memory Latency

Unranked models have fewer than 3 installations. They are listed at their median but receive a rank only once 3 separate installations have published results.

SoAI Bench · scoring

What a score means

A certified SoAI Bench run measures compute throughput, memory bandwidth and dispatch latency on one GPU. The overall score combines compute and memory scores using the weights below. Higher scores are better, including the latency score, which measures dispatch rate and does not contribute to the overall score.

Each installation contributes one score: its median across runs. Rankings use the median of those scores.

Overall = 0.40 × Compute + 0.60 × Memory

For example, 0.40 × 3,334 + 0.60 × 398 = 1,572.4, rounded to 1,572.
Score Measures One point equals Weight
Compute Operations per second across the compute kernels 1 GOPS 40%
Memory Device memory bandwidth 1 GB/s 60%
Latency How quickly the GPU accepts and runs small kernel dispatches 20 dispatches per second Reported only

A certified run

Each certified run starts with at least 120 seconds of warm-up, followed by 5 measured passes. The formula is applied to the median compute throughput and median memory bandwidth of those passes; variance, drift, temperature and power are published beside them as diagnostics.

soai.to recalculates pass scores and medians from the uploaded evidence, then checks publication eligibility before a result appears here.

  1. Warm-up 120 s or more
  2. 1
  3. 2
  4. 3
  5. 4
  6. 5
  7. Median

SoAI Bench · publishing

How a result gets published

Publish results directly from SoAI: review what will be shared, publish it, and get a link. No website account or manual file upload is needed.

  1. Run

    Open Hardware in SoAI, choose a GPU and start a certified run on the standard profile. Quick and stress runs stay in your local history.

  2. Review

    Review exactly what will be published before sending it. Results are shared under the name Anonymous.

  3. Publish

    The run gets a public result page. Open it or copy its link from the result dialog.

Part of a public result
{
  "display_name": "Anonymous",
  "gpu": {
    "gpu_name": "Intel Arc B580",
    "vendor": "Intel",
    "driver_version": "25.05.32567"
  },
  "runtime": {
    "os": "Linux",
    "opencl_platform_name": "Intel(R) OpenCL Graphics",
    "cpu_name": "Intel(R) Core(TM) i7-13700K",
    "system_ram_gb": 30.56,
    "soai_version": "1.2.0"
  },
  "benchmark": {
    "overall_score": 1572,
    "compute_gops": 3333.66,
    "memory_gbs": 397.75,
    "latency_us": 10.86
  }
}

Never published

GPU UUID, PCI address, hostname, IP address and any other local device identifier.

Every run counts

Repeated runs on the same hardware are recorded separately; resending the same publication does not add another run. Each installation contributes one median, so publishing more runs does not give it extra weight in the rankings.

A lasting record

Published scores stay public. Clearing local history in SoAI leaves them unchanged, and moderation removes invalid results.

Download SoAI Hardware & GPU docs