Maximizing Machine Learning Performance with Heterogeneous Computing Resources

Alibaba researcher Zhang Weifeng was invited to speak at the 2017 International Workshop on Mathematical Issues in Information Sciences (MIIS) held from December 14‑17, 2017 at the Chinese University of Hong Kong, Shenzhen, co‑organized by the Shenzhen Big Data Research Institute and the university.

His presentation, titled “Maximizing Machine Learning Performance with Heterogeneous Computing Resources,” explored how to achieve optimal machine‑learning performance by exploiting heterogeneous computing.

Zhang holds a Ph.D. in Computer Science and Engineering from UCSD and has extensive experience in high‑performance multi‑threaded, multi‑core microprocessor architecture, dynamic optimization, and GPU compiler optimization from his work at Qualcomm, Microsoft, and ARM.

He argues that while current AI research focuses on improving algorithmic accuracy, equally important is the use of growing heterogeneous compute power to maximize performance.

The talk emphasized three levels for improving machine‑learning performance: (1) Re‑architecting microprocessors and accelerators (CPU, GPU, FPGA) to meet the inevitable demand for heterogeneous acceleration; (2) Tackling the memory wall through hierarchical memory design, data reuse, compute overlap, and model quantization/compression; (3) Integrated software‑hardware deep optimization, targeting compiler techniques and optimization strategies for heterogeneous architectures.

By addressing compute limitations, Zhang suggests that many approximations and simplifications in AI models can be avoided, enabling breakthroughs that were previously impossible.

The conference theme focused on using modern computing technologies to solve frontier challenges in information science and big‑data analysis, featuring contributions from leading universities such as Stanford, Cornell, UCSD, UCLA, UIUC, Princeton, Waterloo, and industry labs like Microsoft Asia Research, Alibaba Infrastructure R&D, and Tencent AI Lab.