Deep learning accelerators offer tremendous gains in compute throughput. That said, accelerators are only as good as the software compilers built to program them.

I will present VTA, a flexible deep learning accelerator design that is fully integrated and open-sourced with the Apache TVM deep learning compiler stack. I will describe the features and the abstraction layers that compose this complete hardware-software stack. Finally I will discuss how it has enabled researchers to perform full system evaluations of novel research ideas such as numerical system exploration.