Model Accuracy and Performance for INT8 and FP32¶
The following table presents the absolute accuracy drop calculated as the accuracy difference between FP32 and INT8 representations of a model:
| Intel® Core™ i9-12900K @ 3.2 GHz (AVX2) | Intel® Xeon® 6338 @ 2.0 GHz (VNNI) | iGPU Gen12LP (Intel® Core™ i9-12900K @ 3.2 GHz) | |||
|---|---|---|---|---|---|
| OpenVINO Benchmark Model Name |
Dataset | Metric Name | Absolute Accuracy Drop, % | ||
| bert-base-cased | SST-2 | accuracy | 0.11 | 0.34 | 0.46 |
| bert-large-uncased-whole-word-masking-squad-0001 | SQUAD | F1 | 0.87 | 1.11 | 0.70 |
| deeplabv3 | VOC2012 | mean_iou | 0.04 | 0.04 | 0.11 |
| densenet-121 | ImageNet | accuracy@top1 | 0.56 | 0.56 | 0.63 |
| efficientdet-d0 | COCO2017 | coco_precision | 0.63 | 0.62 | 0.45 |
| faster_rcnn_ resnet50_coco |
COCO2017 | coco_ precision |
0.52 | 0.55 | 0.31 |
| resnet-18 | ImageNet | acc@top-1 | 0.16 | 0.16 | 0.16 |
| resnet-50 | ImageNet | acc@top-1 | 0.09 | 0.09 | 0.09 |
| resnet-50-pytorch | ImageNet | acc@top-1 | 0.13 | 0.13 | 0.11 |
| ssd-resnet34-1200 | COCO2017 | COCO mAp | 0.09 | 0.09 | 0.13 |
| unet-camvid-onnx-0001 | CamVid | mean_iou@mean | 0.56 | 0.56 | 0.60 |
| yolo-v3-tiny | COCO2017 | COCO mAp | 0.12 | 0.12 | 0.17 |
| yolo_v4 | COCO2017 | COCO mAp | 0.52 | 0.52 | 0.54 |