Hamidreza Khazaei

Ai Hw Sw Co-design @Meta

Menlo Park, CA, US

I co-design hardware and software for smart glasses, taking workloads like graphics, neural networks, computer vision, and audio and optimizing them on DSPs—tuning both SW and HW to improve area, power, and performance- Led HW/SW co-design and profiling of a core vision workload on DSP, enabling migration from application processor to DSP and driving power savings while supporting a more compute-heavy configuration; delivered 30% cycle reduction and 20% memory reduction via low-level optimizations- Co-designed and optimized ML kernels for Cadence HiFi1/HiFi5 DSPs (including TIEs and pipeline changes), providing architecture feedback that reduced DSP area and power while preserving required ML performance- Implemented and validated production ML operators (quantized convolution, quantized linear, etc.), improving MAC utilization to 83%- Prototyped running parts of the graphics/texture pipeline on DSP (decompression, format conversion), achieving sub-1 cycle per pixel in key conversions and laying groundwork to future-proof graphics offload on DSP- Evaluated and enabled a vision neural network on DSP, using roofline and per-layer analysis to demonstrate ~1 ms inference latency, unlocking a low-power, privacy-preserving on-device tracking solution.

Stanford University

Masters of Science, Electrical Engineering

University of Toronto

Bachelors of Applied Science, Electrical Engineering