Portrait of Hossein Askari

Hossein Askari

Contact:

m.h.askari.hemmat (AT) gmail (dot) com

I am a Research Engineer at Meta, where I work on developing new methods for efficient AI models. I received my Ph.D. in Electrical Engineering from Polytechnique Montréal.

My research focuses on efficient deep learning: quantization, ML performance modeling, and hardware accelerators for neural networks, spanning algorithms, compilers/runtimes, and custom RISC-V hardware.

Selected Publications

LEAP hardware ranking
LEAP predicts each training job's throughput on every candidate hardware type at submission time and ranks them by throughput per unit cost, routing tens of thousands of jobs a day across Meta's training fleet and saving an estimated 3% of training compute.
arXiv 2026 / PDF
QGen loss landscape
A theoretical model of quantization as regularization with an approximate generalization bound, validated on 2000+ CNN and transformer models on CIFAR-10/100 and ImageNet.
arXiv 2024 / PDF
Quark
An integer RISC-V vector processor in GF 22FDX for sub-byte quantized neural network inference.
ISCAS 2023 / arXiv / Code
BARVINN logo
An open-source, arbitrary-precision FPGA DNN accelerator controlled by a RISC-V CPU.
ASP-DAC 2023 / arXiv / Code / Docs
U-Net quantization
A flexible fixed-point quantization method for U-Net, evaluated on three medical image segmentation datasets.
MICCAI 2019 / arXiv / Code / Slides

Talks & Workshops

  • [2021/05] BARVINN: Barrel RISC-V Neural Network Accelerator @ CMC Accelerating AI 2021 (Event / Video)
  • [2020/03] Hardware Aware Acceleration for Deep Neural Networks @ CMC Workshop, Montréal (Event / Slides)
  • Workshop on New Methods on Designing Digital Systems: open-source design tools, RISC-V, and Chisel (Website / Slides)