← sources

We measured 602 vision models across 5 different GPUs

source: reddit (reddit:computervision)published 2026-11-11 18:04ingested 2026-11-12 02:38

retained text

We measured 602 timm ImageNet classifiers with TensorRT INT8 on five NVIDIA GPUs: L4, RTX 2080 Ti, Jetson AGX Thor, AGX Orin and Orin NX. Each line in the chart is a GPU's Pareto front.

The best model depends on the GPU. Of the 149 models measured on all five GPUs, only 2 are on every front.

* PiT-B is the best choice on Thor (83.9% in 1.05 ms), but 2.3-2.8× slower than the front everywhere else.
* ConvNeXt-B is on the front everywhere except the L4.
* MobileNetV2-1.4 is on the front everywhere except Thor.  

INT8 cost: a median of −0.76 top-1 against timm's FP32 numbers, across 609 models.

Method: TensorRT, batch 1, GPU compute only, timm default input size. Accuracy is measured on the full ImageNet val set.   
  
Every model has its own page per GPU, with full numbers: [https://embedl.com](https://embedl.com)  
  
What should be measured next?

document fetched 2026-11-12 00:24 · canonical https://reddit.com/r/computervision/comments/1xtow4z/we_measured_602_vision_models_across_5_different