Extension projects to enable collaborative benchmarking, design space exploration and optimization of ML and AI Systems

[gfursin]

No description provided.

[gfursin]

• Test that MLPerf inference benchmark works using the CM interface across different implementations, models, frameworks, data sets and platforms. See the CM interface with available variations and flags here.

  • Increase the coverage of the reference implementation of the MLPerf inference benchmark. See the current status here.
  • Add more ML models (Hugging Face, Nvidia benchmark, papers) compatible with the MLPerf inference benchmark as CM scripts, test the benchmark and add models as variations here.
  • Test TFLite C++ implementation of the MLPerf inference benchmark and extend coverage to other models and platforms. Attempt to reproduce past MLperf submissions with this new interface: (data center and edge)
    • Extend MLPerf CM script to reproduce past open division submissions from MLPerf inference v2.1 (data center and edge) and prepare new submissions using the latest versions of all dependencies
  • Test C++ implenentation of the MLPerf inference benchmark and increase the coverage
  • Test DeepSparse implementation
  • Compare different implementations and attempt to optimize them using different batch sizes, thread numbers, etc
  • Participate in MLPerf inference v3.0 submission
  • Improve tutorials and documentation to make it easier for the community to understand CK2 (CM) concepts, use CM interface for unified ML Systems benchmarking and extend CM automations!

• Some performance/accuracy experiments per model

  • ResNet50
    • Integrate tflite cpp code to generic-cpp code, make all scenarios run (currently only singlestream works)
    • Try cuda device for tflite
    • Add new models which can work with imagenet dataset to open division
    • Check performance of quantized models on different backends
    • Compare performance of reference and Nvidia implementation on GPUs
    • Compare performance of reference and Intel implementation on CPUs
    • Does TVM improve performance of any model/system.scenario?
  • Bert
    • Try different bert models trained on squad
    • Try quantized models on different backends
    • Compare performance of reference and Nvidia implementation on GPUs
    • Compare performance of reference and Intel implementation on CPUs
    • Does TVM improve performance of any model/system.scenario?
  • RetinaNet
    • Dry different retinanet models trained on openimages
    • Try quantized models on different backends
    • Try running the NMS part on CPU and rest on GPU
    • Compare performance of reference and Nvidia implementation on GPUs
    • Compare performance of reference and Intel implementation on CPUs
    • Does TVM improve performance of any model/system.scenario?

• Test/improve the CM script for the light MLPerf inference benchmark to benchmark and optimize any ONNX model with loadgen but without data sets and accuracy!

• Test/improve CM interface to automatically prepare and run TinyMLPerf and prepare official tutorial. Use OctoML's submission as a starting point: results, code

  • Try another device such as Arduino Nano 33 BLE sense if supported in MLPerf
  • Participate in TinyMLPerf inference v3.0 submission

• Test and improve individual CM scripts across different software/stacks to be reused in any R&D project - will be useful for our reproducibility initiatives and artifact evaluation at conferences

[arjunsuresh]

Reference for adding CUDA for tflite-cpp