Two devices side by side, measured the same way: the same PyTorch models on the stock, publicly released stack, each at its own saturation point. Every comparison shows the overall ratio, the Training, Inference and Compute ratios, and every model and precision underneath.
Each figure is the first device's overall score divided by the second's.
Each figure is the first device's overall score divided by the second's.
Each figure is the first device's overall score divided by the second's.
Each figure is the first device's overall score divided by the second's.
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).