Hello! I am Yang Haoxiang (prounce: “Young-How-Hsiang” or “Yeung-Holl-Tseung”). I am passionate about technology and innovation. Currently, I am focusing on learning and researching high-performance computing and computer architecture, particularly interested in AI Infrastructure.
September 2024 - June 2028
Guangzhou, Guangdong (Canton), China
September 2024 - June 2028
🏆 6th place overall. \n By 1. rewriting part of MPI with NCCL and fine-tuning to achieve cross-GPU asynchronous communication for cudaManagedMemory; 2. kernel fusion; 3. reducing redundant computations, we achieved a 1.48x speedup on a cluster of 16×A800, which is several times faster than the CPU-only OpenFOAM.\n Implemented data parallelism for llama.cpp.
Ranked 28th, did not advance to the finals. Optimized the single‑concurrency throughput of Qwen3.5 on a single Sugon DCU(A AMD’s GPU-like compute card). Using various optimization techniques, achieved an average 1.93x throughput improvement across different batch sizes compared to the baseline.
Participated as an observer and learned about FFTW optimization on an advanced domestic(in China) architecture.
🏆 3rd place in the finals. Optimized the CESM software on a domestic(in China) ARM many‑core cluster. Ran ML models for predicting marine chlorophyll on the same cluster.
Turbofold is an efficient solution for running Alphafold3 inference on distributed clusters. My main contribution was the design of a sequence‑parallel‑like scheme for the Diffusion Module, and I also participated in the design of efficient Allgather and Alltoall communication.\n The former refers to applying a sequence‑parallel approach to operations like attention on a protein sequence in AF3, similar to LLM inference, which helps with long sequences and low sampling (similar to concurrency in LLMs). The work has not been published.
🏆 NOIP 2021 Provincial First Prize, WC 2022 Silver Award
