Helprack has been helping lower code from ONNX dialect of MLIR to a proprietary custom dialect that is targeted at large scale acceleration of ML training workloads for a Bay Area based Enterprise

This customer has developed an AI/ML chip that has a compute fabric akin to a GPU and can compute on the Edge. The goal is to translate all 192 ONNX operations to the custom dialect of MLIR.

Helprack delivered the first 50 ONNX operations that target the RESNET50 network to custom dialect. This is progressively lowered to processor specific dialect of MLIR and then to native instructions. The RESNET50 model has been shown to run on the custom processor.