SPRADB8 may   2023 AM62A3 , AM62A7

 

  1.   1
  2.   Abstract
  3.   Trademarks
  4. 1Introduction
  5. 2AM62A Processor
  6. 3Deep Learning Benchmarks
  7. 4Retail Checkout Scanner Application
  8. 5Core Loading
  9. 6Part Selection
  10. 7Power Usage
  11. 8Summary
  12. 9References

3 Deep Learning Benchmarks

Table 3-1 shows benchmarks on the AM62A running the Linux Edge AI SDK version 8.6 on the Starter Kit EVM revision E2. Note that these numbers reflect 1.7 TOPS of maximum performance, as the E2 EVM’s PMIC supplies 0.75 VDD core voltage. Achieving the full 2 TOPS requires 0.85 VDD. EVMs after E2 will use the updated PMIC for full performance entitlement.

Table 3-1 Benchmarks on the AM62A
Model Name Accuracy (C7xMMA) Frames-per-Second (C7xMMA) Frames-per-Second (CPU) Resolution
Classification Network – Accuracy is Top1 Metric
TFL-CL-0010-mobileNetV2 74.51 251 5.5 224 x 224
Object Detection Network – Accuracy is mAP50-95
TFL-OD-2000-ssd-mobV1-coco-mlperf-300x300 26.14 152 2.4 300 x 300

TI’s model zoo hosts many more models that are pretrained and fully accelerated on the C7xMMA. The most up-to-date benchmarks (relative to software and EVM version) can be found in TI’s Model Analyzer tool [1].