NVIDIA GPU throughput by precision
Dense, not sparseNVIDIA's big numbers usually assume “2:4 sparsity” (half of every four weights are zero), which doubles the figure but which almost no real model uses. This page shows dense numbers unless you switch to sparse.
Headline numbers mix precisionsThe single “Tensor TFLOPS” or “AI TOPS” figure on a spec sheet is FP16 for Ampere, FP8 for Ada and FP4 for Blackwell, all with sparsity. Compare one column at a time instead.
GeForce runs half speed in FP32 accumulateOn GeForce cards, FP16, BF16, TF32 and FP8 matrix math runs at half speed when results are summed in FP32, which is what PyTorch does. Workstation and data-centre cards don't have this limit. The FP16 and BF16 columns show the FP32-accumulate speed.
Real cards run below these peaksThese are peaks at the rated boost clock. Under heavy matrix math cards hit their power limit and clock down: the four cards benchmarked below sustained 74–82% of their BF16/FP16 peaks.
Every card
Click a column to sort; click a card for its sources and the maths. Plain numbers are NVIDIA's own; dotted ones are derived (hover for how); — means no hardware support.
Measured on real cards
Where the numbers come from
Dense ops per clock per SM, by architecture
From NVIDIA's whitepapers. Multiply by SM count and clock to get FLOPS. “GF” = GeForce.
Checks
Downloads
All data: gpu-flops.json · gpu-flops.csv · benchmark script gpubench.py · raw benchmark logs bench/