Leaderboard Β· All comparisons Β· LynxBenchAI 26Q3 Index
Device pages: NVIDIA RTX 6000 Ada Generation Β· NVIDIA GeForce RTX 4090
On the LynxBenchAI 26Q3 Index, the NVIDIA RTX 6000 Ada Generation scores 1.02x the NVIDIA GeForce RTX 4090 overall (Training 1.11x, Inference 0.91x, Compute 1.08x). It leads on Training and Compute and trails on Inference.
| Score | Ratio | RTX 6000 Ada Generation | RTX 4090 |
|---|---|---|---|
| Overall (GT) | 1.02x | 457.36 | 450.05 |
| Training | 1.11x | 373.92 | 335.94 |
| Inference | 0.91x | 597.42 | 659.37 |
| Compute | 1.08x | 396.70 | 367.89 |
From the LynxBenchAI 26Q3 Index category scores, which are not raw throughput; a lead under 10% counts as level.
| Model | Precision | Ratio | RTX 6000 Ada Generation | RTX 4090 |
|---|---|---|---|---|
| compute | ||||
| Dense MatMul | BF16 | 1.05x | 298.13 | 283.79 |
| Dense MatMul | FP16 | 0.83x | 245.50 | 295.41 |
| Dense MatMul | FP32 | 1.49x | 356.80 | 239.94 |
| Dense MatMul | FP64 | 1.01x | 9.55 | 9.48 |
| Sparse MatMul | FP32 | 1.16x | 696.19 | 598.47 |
| Sparse MatMul | FP64 | 0.97x | 645.83 | 666.20 |
| inference | ||||
| DistilBERT | FP16 | 0.83x | 886.10 | 1069.74 |
| DistilBERT | FP8 | 0.86x | 577.39 | 673.12 |
| DistilBERT | INT8 | 0.95x | 530.42 | 561.18 |
| DistilGPT2 | FP16 | 0.84x | 801.58 | 957.02 |
| DistilGPT2 | FP8 | 0.82x | 506.53 | 617.67 |
| DistilGPT2 | INT8 | 0.84x | 483.17 | 577.05 |
| Flan-T5 Small | FP16 | 0.79x | 1049.28 | 1333.43 |
| Flan-T5 Small | FP8 | 0.95x | 740.46 | 778.03 |
| Flan-T5 Small | INT8 | 1.01x | 657.38 | 650.15 |
| GraphGPS (Peptides) | FP16 | 0.91x | 318.37 | 351.19 |
| GraphGPS (Peptides) | FP8 | 0.96x | 167.34 | 173.88 |
| GraphGPS (Peptides) | INT8 | 0.94x | 157.35 | 167.13 |
| MobileNetV3 | FP16 | 1.02x | 1547.15 | 1519.03 |
| MobileNetV3 | INT8 | 1.07x | 467.17 | 437.31 |
| training | ||||
| DistilBERT | BF16 | 0.96x | 455.59 | 472.39 |
| DistilBERT | FP32 | 1.02x | 636.08 | 623.27 |
| DistilGPT2 | BF16 | 0.96x | 262.78 | 272.36 |
| DistilGPT2 | FP32 | 1.01x | 476.79 | 469.76 |
| Flan-T5 Small | BF16 | 1.06x | 292.43 | 276.10 |
| Flan-T5 Small | FP32 | 1.16x | 586.86 | 506.61 |
| GraphGPS (Peptides) | BF16 | 1.42x | 214.53 | 151.20 |
| GraphGPS (Peptides) | FP32 | 1.54x | 243.50 | 158.56 |
| MobileNetV3 | BF16 | 1.04x | 280.24 | 268.26 |
| MobileNetV3 | FP32 | 1.08x | 424.68 | 394.16 |
Each figure is the first device's overall score divided by the second's.
Measure your own machine: pip install lynxbench-ai Β· Install guide β
A LynxBenchAI score is sustained throughput on a fixed catalogue of PyTorch models, measured after a warm-up at a batch size raised until throughput stops improving, so it is what the machine holds under load, not its best instant. Training, Inference and Compute are three separate scores; GT combines them (Training and Inference weigh 0.4 each, Compute 0.2).
Every run uses the stock, publicly released PyTorch stack, so a result is broadly indicative of the out-of-the-box PyTorch experience on that hardware, and so of custom PyTorch models too. It does not run your own model, and it is not a vendor-tuned peak like an MLPerf submission.
Scores compare only within one Index: the 26Q3 Index fixes the models, precisions, correctness thresholds and scoring, and a later Index may change them. Each test is one timed window, the scores have no 0-100 scale, and a precision a device cannot run scores zero rather than being skipped.
LynxBenchAI is built by TechnoLynx, which the European Commission engaged as an independent expert advisor on AI performance and benchmarking for AI Gigafactories under the EuroHPC Joint Undertaking (case study).
The Personal Edition submits your benchmark results to TechnoLynx's servers as part of participating in the public leaderboard. If that isn't acceptable for your use case, contact TechnoLynx about Press, Pro, or Enterprise licenses.
LynxBenchAI End-User License Agreement · LynxBenchAI Privacy Policy