Performance of TC

TC can generate competitive code in a variety of cases thanks to its Autotuner (see our companion paper: arXiv). We will provide a set of benchmarks to illustrate the cases in which it is recommended to use TC.

As a general rule of thumb, TC is a good candidate to rapidly prototype new ML layers and integrate them without writing a single line of CUDA code. For existing, computation bound layers, it should be expected that TC performance will not beat libraries such as CUBLAS and CUDNN except in very specific corner cases, described in our paper.

For the cases where efficient library implementations exist (e.g. matmul, convolutions), it is usually recommended to use existing libraries, for now.