NVIDIA · GPU · LynxBenchAI 26Q3 Index, recorded 2026-10-22
On the LynxBenchAI 26Q3 Index, the NVIDIA L4 ranks 31 of 46 devices overall with a GT of 150.07 (Training 116.66, rank 33; Inference 209.08, rank 33; Compute 125.79, rank 24); overall it scores 0.16x the NVIDIA H100 80GB HBM3 and 0.33x the NVIDIA GeForce RTX 4090.
Rank among all 46 devices; further right is higher.
| Compared with | Overall | Training | Inference | Compute |
|---|---|---|---|---|
| NVIDIA H100 80GB HBM3 | 0.16x | 0.16x | 0.12x | 0.26x |
| NVIDIA GeForce RTX 4090 | 0.33x | 0.35x | 0.32x | 0.34x |
| Score | Next above | Ratio | Next below | Ratio |
|---|---|---|---|---|
| Training | NVIDIA GeForce RTX 5060 | 0.97x | NVIDIA GeForce RTX 4060 Ti | 1.26x |
| Inference | NVIDIA GeForce RTX 5060 | 0.94x | DGX Spark (NVIDIA GB10) | 1.15x |
| Compute | NVIDIA GeForce RTX 5080 | 0.94x | NVIDIA GeForce RTX 5070 Ti | 1.08x |
| Model | Precision | Sustained |
|---|---|---|
| MobileNetV3 | FP32 | 144.59 |
| MobileNetV3 | BF16 | 101.95 |
| Flan-T5 Small | FP32 | 166.21 |
| Flan-T5 Small | BF16 | 89.91 |
| DistilGPT2 | FP32 | 137.30 |
| DistilGPT2 | BF16 | 75.36 |
| DistilBERT | FP32 | 189.33 |
| DistilBERT | BF16 | 141.94 |
| GraphGPS (Peptides) | FP32 | 68.68 |
| GraphGPS (Peptides) | BF16 | 84.39 |
| Model | Precision | Sustained |
|---|---|---|
| MobileNetV3 | FP16 | 534.89 |
| MobileNetV3 | INT8 | 146.74 |
| Flan-T5 Small | FP16 | 408.91 |
| Flan-T5 Small | INT8 | 202.53 |
| Flan-T5 Small | FP8 | 275.49 |
| DistilGPT2 | FP16 | 304.72 |
| DistilGPT2 | INT8 | 164.24 |
| DistilGPT2 | FP8 | 183.47 |
| DistilBERT | FP16 | 358.58 |
| DistilBERT | INT8 | 184.25 |
| DistilBERT | FP8 | 215.96 |
| GraphGPS (Peptides) | FP16 | 96.20 |
| GraphGPS (Peptides) | INT8 | 50.07 |
| GraphGPS (Peptides) | FP8 | 46.91 |
| Model | Precision | Sustained |
|---|---|---|
| Dense MatMul | FP32 | 96.43 |
| Dense MatMul | FP16 | 103.94 |
| Dense MatMul | BF16 | 111.93 |
| Dense MatMul | FP64 | 3.24 |
| Sparse MatMul | FP32 | 211.34 |
| Sparse MatMul | FP64 | 181.54 |
LynxBenchAI 26Q3 Index · GT = LynxBenchAI Global Score. Every precision result is sustained throughput under a declared, bounded optimisation budget — no collapsed single number below the category level.
Each figure is the first device's overall score divided by the second's.
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).
Measure your own NVIDIA L4: pip install lynxbench-ai · Install guide →
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