I analyze deep learning applications (CNNs, LLMs, GNNs) and create optimal SW mappings to current and future GPU architectures. I find the limiting factors and develop new GPU HW features to breakthrough them. I developed the high level HW architecture for XMX (Intel's matrix accelerator for GPUs), specified the XMX ISA, developed the performance model, and wrote the first convolution and matrix multiply kernels for XMX. I analyzed and modeled a variety of GPU architecture features for deep leaning including structured sparsity, unstructured sparsity, cache policies, pre-fetching, compression, low precision arithmetic, and interconnects.

Developed proposals and prototypes for changes to the OpenCL standard to support Intel's new GPU features, such as shared virtual memory. Analyzed proposals from industry participants to ensure they were efficiently implementable on Intel's current and future GPUs

Worked in developer relations optimizing game engines for Intel CPUs and GPUs.Developed automatic offline 3D mesh simplification system along with a continuously variable LOD runtime system. My code shiped in Shockwave 3D, 3D Studio Max, and a few games.

Implemented a compiler for a Domain Specific Language (Adapt) for image processing on iWarp (a systolic array based supercomputer).