DAIBench (DiDi AI Benchmarks) aims to provide a set of AI evaluation sets for production environments, spanning different types of GPU servers and cloud environments, to provide users with effective and credible test results for future hardware selection , software and library optimization, business model improvement, link stress testing and other stages to lay a solid data foundation and technical reference.
- Layerwised benchmarking, from hardwares(L1), operators(L2) to models(L3), higher level benchmarking is TBD.
- Cloud-native benchmarking, containerized deploying, easy to use.
- Multi-cloud benchmarking, results are useful for price/performance considerations.
DAIBench comprehensively considers the existing GPU performance testing tools, and divides the indicators into hardware layer, framework (operator) layer, and algorithm layer.
For each level, DAIBench currently supports the following tests:
| Layer | Supported Test |
|---|---|
| Hardware layer | Focusing on the indicators of the hardware itself, such as peak computing throughput (TFLOPS/TOPS) calculation indicators and memory access bandwidth, PCIe communication bandwidth and other I/O indicators. |
| Frame/operator layer | Evaluating the computing power of commonly used operators (convolution, Softmax, matrix multiplication, etc.) based on mainstream AI frameworks. |
| Model layer | Performing end-to-end evaluation by selecting models in a series of production tasks. |
cd <test_folder>
bash install.sh
bash run.sh
For GPU test, please install suitable nvidia-driver and cuda first.
Current operator layer is using DeepBench
cd operator
bash install.sh # download source code & prepare nccl
To run GEMM, convolution, recurrent op and sparse GEMM benchmarks:
bin/gemm_bench <inference|train> <int8|float|half>
To execute the NCCL single All-Reduce benchmark:
bin/nccl_single_all_reduce <num_gpus>
The NCCL MPI All-Reduce benchmark can be run using mpirun as shown below:
mpirun -np <num_ranks> bin/nccl_mpi_all_reduce
num_ranks cannot be greater than the number of GPUs in the system.
docker and nvidia-docker is required for model testing. To run specific model, please read Readme.md in the folder.
General test procedure:
- Download dataset
- Preprocess dataset (if needed)
- Build docker
- Launch benchmark
- Get result
See wiki for guidelines.
Welcome to contribute by creating issues or sending pull requests. See Contributing Guide for guidelines.
DAIBench is licensed under the Apache License 2.0.