Siyuan Zhang

About Me

I am a graduate student from Shanghaitech University. I am looking for PhD position in 24 Spring and 24 Fall in the field of HPC and MLsys.

Research Interests

• HPC : MPI, GPU, HPC4AI
• MLSys

Experience

Research Assistant @ L.I.O.N. & NeuDim

• Co-author NEPHELE: A Neural Platform for Highly Realistic Cloud Radiance Rendering, as the second author.

  • Reduce hash grid size by 8.7 times by introducing Perfect Spatial Hashing into NGP.
  • Develop a multi-GPU, multi-node extrem low latency NGP rendering framework, achieving 50 FPS @ 1080 resolution for 12 NGP on 4 GPUs in one node and 37 FPS @ 1080 resolution for 12 NGP on 4 GPUs in 2 nodes using CUDA awared MPI.

• Research in workloads of DNN Training Jobs in HPC

Student Cluster Competition (SCC) @ Geekpie_HPC Shanghaitech University

High Performance Computing (HPC) related research. Continually preparing and aiming for the top-tier student supercomputing competitions, including ASC, ISC and SC. Mainly in charge of HPC tasks related with AI, MPI and CUDA.

• ASC22 @ USTC :
  • Manage hardwares under restricted power constraints (3KW, 3 nodes, 8 GPUS).
  • Use DeepSpeed to accelerate of YUAN A Large Language Model Pretrain Model.
  • Fine-tuned HPC benchmarks HPCG and HPL.
• ISC22 @ Online :

• Profile and optimize ICON a MPI Program for earth system model (3rd place).

• SC21 @ Online (2nd place):
  • Reproduce the correctness of a SC research ramBLe (1st place).
  • Run and tuning a multi-node multi-GPU DNN image segmentation application CosmicTagger. (2nd place).
• ISC21 @ Online:
  • Modify a GPU optimized HPCC benchmark
  • Research on MPI and OpenMP balance for WRF
  • Contributed for collective_profiler a set of tools and libraries to profile Alltoallv MPI calls.

Teaching Assistant

CS171: Computer Graphics ShanghaiTech University

Skills

Programming Languages

C/C++, CUDA, Python

Frameworks, Libs

OpenMP, MPI, Pytorch, OpenGL, OpenCV, NCCL

Tools

Nsight System, Nsight Compute, Intel Vtune, Intel Trace Analyser and Collector, IPM.

Publications

Powered by Jekyll and Minimal Light theme.