[antinucleon]

Summary: Tensor program is able to be optimized by using machine learning and transfer learning. The numerical program optimization model is trained on feature from low-level AST of the program. Experiments: Tasks: ResNet, MobileNet, LSTM LM, DQN Hardware: CUDA/ARM GPU/ARM CPU Speed up compare to CUDNN, TensorFlow Lite and ARMComputeLib: ~from 1.2X to 3.8X faster in end-to-end test.

[namibj]

Be careful what they benchmark against. The Maxwell/Pascal GEMM routines are _very_ optimized, since Nervana Systems (in their infancy) made an assembler for them which they showed off/demonstrated by implementing GEMM (single-precision dense matrix*matrix multiplication). It was lacking like 2% from the theoretical performance, as predicted by the clockrate, but they had no idea why these 2% were missing. They hand-scheduled the instructions to get around the limited number of register banks, and the associated bank conflicts, a concept normal CPUs afaik only see with DRAM. At that time, the official GEMM, albeit hand-optimized, only reached like 80% of the theoretical value.

[crowwork]

the benchmarks are not about GEMM, but real-world deep learning workloads which could have very different characteristics from GEMM

[namibj]

I just wanted to caution that one has to be careful what one is comparing against, as the libraries got significant speed improvements over time, without that being widely advertised. So it matters a lot if one compares this library against CuDNN from 1 month ago, or to CuDNN from 2 years ago. The latter is _much_ slower.

The GEMM example was just there as the details of the optimization have been published, unlike most other hand-tuned assembler routines for DNN workloads.

[antinucleon]

CuDNN v7 was used in the experiments, in the experiments parts each comparison was listed with version or commit number.

[namibj]

Well, I didn't RTFA. This was not meant to be specific for this article though.