Hello! 👋
I'm Yuwei (Peter) Sun
MCS @ UIUC · High-Performance Computing Engineer
News
These updates mark key milestones along my journey.
Swipe or scroll horizontally
最新动态
这些更新记录了我旅程中的重要节点。
左右滑动或横向滚动查看更多
About Me
I build high-performance computation systems by applying HPC principles to model inference and infrastructure. My goal is to make large-scale AI and HPC applications faster, cheaper, and smarter.
Keywords
Skills
关于我
我通过将 HPC 原理应用到模型推理和基础设施，构建高性能计算系统，让大规模 AI 与 HPC 应用更快、更高效、更智能。
关键词
技能
Work Experience
Software Engineering Intern
Waymo
- Built an end-to-end ML workflow for agent-handover prediction, covering data sampling, model training, offline evaluation, and inference.
- Integrated inference scores into driving-set candidate selection to support log-based evaluation without requiring simulation-derived queries.
High Performance Computing Intern
Flatiron Institute, Simons Foundation
- Extended nifty-ls to support multiple sinusoidal basis terms for flexible Lomb-Scargle periodograms.
- Re-engineered compute-heavy Python routines as C++ extensions via nanobind and FINUFFT for ~100× speedup, then reached ~300× with 16-thread OpenMP.
- Ported FastChi2 assembly and periodogram post-processing to NVIDIA A100 GPUs with CuPy, achieving ~50× acceleration over single-threaded C++ and up to ~5,600× over the original Python implementation.
- Reduced FastChi2 99th-percentile numerical error from 1e-3 to 1e-6 by rewriting trigonometric summation and replacing NumPy FFT with FINUFFT.
Part-time Research Engineer
New York University
- Benchmarked and compiled Gaussian and Amber on NVIDIA Grace Hopper ARM CPUs, evaluating energy efficiency and compatibility.
- Led migration from a physical HPC cluster to an OpenShift-based on-prem cloud using Podman, Singularity, and Kubernetes.
HPC Assistant
NYU High-Performance Computing Team
- Enhanced Reform with sequential processing, compressed file handling, and parallel execution for genomics workloads.
- Updated ReformWeb to run Reform on HPC servers via a web UI with Flask, SQLite, Redis, Werkzeug, and Jinja2.
- Built CI/CD with GitHub Actions and Python unittest plus a Bash-based verification pipeline with TTL-based log retention.
Part-time Research Engineer
NYU High-Performance Computing Team
- Translated CUDA codebases to HIP with HIPify, enabling AMD GPU support for AlphaFold and DualPhysics on ROCm.
- Contributed fixes to HIPify for Math_constant.h issues to improve CUDA-to-HIP translation reliability.
- Evaluated cloud based ARM CPU suitability on AWS Graviton and GCP Axion by compiling HPL, Amber24, and Gaussian and benchmarking price-performance.
工作经历
软件工程实习生
Waymo
- 围绕 agent handover 预测，构建覆盖数据采样、模型训练、离线评估和推理的机器学习流程。
- 将模型推理分数接入 driving-set 候选样本筛选，支持基于日志的评估流程，减少对仿真查询的依赖。
高性能计算实习生
Flatiron Institute, Simons Foundation
- 扩展 nifty-ls，使其支持多组正弦基函数，灵活计算 Lomb-Scargle 周期图。
- 通过 nanobind 和 FINUFFT 将计算密集的 Python 例程重构为 C++ 扩展，速度提升约 100 倍；使用 16 线程 OpenMP 后达到约 300 倍。
- 借助 CuPy 将 FastChi2 矩阵组装和周期图后处理迁移至 NVIDIA A100 GPU，相比单线程 C++ 加速约 50 倍，相比原始 Python 实现最高加速约 5,600 倍。
- 重写三角函数求和并以 FINUFFT 替换 NumPy FFT，将 FastChi2 数值误差的第 99 百分位值从 1e-3 降至 1e-6。
研究工程师（兼职）
New York University
- 在 NVIDIA Grace Hopper ARM CPU 上编译并基准测试 Gaussian 与 Amber，评估ARM CPU能效与ARM 生态对于HPC的兼容性。
- 使用 Podman、Singularity、Kubernetes 将物理 HPC 集群迁移计算环境和数据到基于 OpenShift 的本地云。
HPC 助理
NYU High-Performance Computing Team
- 为 Reform 增强顺序处理、压缩文件处理和并行执行，适配基因组工作负载。
- 更新 ReformWeb，使 Reform 可通过 Flask、SQLite、Redis、Werkzeug、Jinja2 在 HPC 服务器上运行。
- 构建 GitHub Actions + Python unittest 的 CI/CD，并用 Bash 验证流水线与 TTL 日志保留。
兼职研究工程师
NYU High-Performance Computing Team
- 用 HIPify 将 CUDA 代码翻译为 HIP，使 AlphaFold 和 DualPhysics 支持 AMD GPU 的 ROCm 平台。
- 为 HIPify 修复 Math_constant.h 相关问题，提高 CUDA 到 HIP 的转换可靠性。
- 在 AWS Graviton 与 GCP Axion 上编译 HPL、Amber24、Gaussian 并基准测试，评估基于 ARM CPU 云计算的性价比。
Projects
Kernel-Level Optimization for GPT-2 (Sep 2025 - Dec 2025)
Optimized GPT-2 inference kernels on A40 GPUs using Tensor Cores, Split-K, and FlashAttention. Reduced MatMul kernel runtime by 94.6% and improved throughput by ~180%; fused attention cut global memory traffic by 85% and global-to-core data movement from ~800 KB to 37–56 KB. Guided kernel tuning with Nsight Systems and Nsight Compute profiling.
Distributed Order System Refactor
Scaled a high-traffic order service from 1,000 to 50,000 QPS using Redis and RocketMQ with MySQL primary-replica replication. Refactored the monolith using DDD and SOLID with JOOQ and Flyway; implemented TCC inventory control with Redis/Lua, asynchronous order creation, delayed payment checks, and optimistic locking for transactional correctness.
Computer Architecture & System Programming Projects
Developed a Python E20 assembler and C++ CPU and cache simulators to model instruction execution and memory behavior.
项目
GPT-2 内核级优化（2025 年 9 月 - 2025 年 12 月）
在 A40 GPU 上通过 Tensor Cores、Split-K 和 FlashAttention 优化 GPT-2 推理内核。相比 CUDA 基线将 MatMul 内核运行时间降低 94.6%、计算吞吐提升约 180%；融合注意力计算使全局内存流量降低 85%，并将全局内存到计算核心的数据搬运量从约 800 KB 降至 37–56 KB。利用 Nsight Systems 和 Nsight Compute 的分析结果指导优化。
分布式订单系统重构
在 MySQL 主从架构中引入 Redis 和 RocketMQ，将订单服务吞吐从 1,000 提升至 50,000 QPS。遵循 DDD 与 SOLID 原则拆分单体系统，并采用 JOOQ 和 Flyway 管理数据访问与迁移；结合 Redis/Lua 的 TCC 库存控制、异步建单、延迟支付核查及乐观锁保障事务正确性。
计算机体系结构与系统编程项目
用 Python 开发 E20 汇编器，并用 C++ 编写 CPU 与缓存模拟器，模拟指令执行与内存行为。
Research
I research heterogeneous computing for scientific applications and wafer-scale GPU architectures. My work includes collaborating with the Flatiron Institute on parallel optimization of periodograms and with the Scalable Architecture Lab on profiling and memory systems for large accelerators.
R01
Branch Prediction Research
ISCA 2026 CBP-NG Workshop
2026
Contributed to RABT: Run-Ahead Block TAGE, accepted to the CBP-NG workshop at ISCA 2026.
Publication: RABT: Run-Ahead Block TAGE
R02
HPC Scientific Software
Flatiron Institute, Simons Foundation
Sep 2025
Poster on optimizing and extending Lomb-Scargle periodogram computations for astronomy using OpenMP and CUDA
Mentor: Dr. Lehman Garrison
Publication: Optimizing and Extending Periodogram Computations for Astronomy
R03
Wafer-scale GPU Architecture
Scalable Architecture Lab
March 2025
Hierarchical distributed page address translation for wafer-scale GPUs to improve memory-system scalability.
Mentor: Prof. Yifan Sun
Publication: HDPAT: Hierarchical Distributed Page Address Translation for Wafer-Scale GPUs
R04
GPU Performance Tooling
Scalable Architecture Lab
Sep 2024
Dynamic binary instrumentation for AMD GPUs to study performance of large-scale accelerators.
Mentor: Prof. Yifan Sun
Publication: Luthier: A Dynamic Binary Instrumentation Framework Targeting AMD GPUs
研究
我研究面向科学应用的异构计算与晶圆级 GPU 架构。工作包括与 Flatiron Institute 合作，优化周期图的并行计算，以及在可扩展体系结构实验室进行大规模加速器的性能分析与内存系统研究。
R01
分支预测研究
ISCA 2026 CBP-NG 研讨会
2026 年
参与 RABT: Run-Ahead Block TAGE 研究，论文被 ISCA 2026 CBP-NG 研讨会接收。
R02
HPC 科学软件
Flatiron Institute, Simons Foundation
2025 年 9 月
海报展示：通过 OpenMP 和 CUDA 优化并扩展 Lomb-Scargle 周期图计算。
导师: Dr. Lehman Garrison
论文: Optimizing and Extending Periodogram Computations for Astronomy
R03
晶圆级 GPU 架构
Scalable Architecture Lab
2025 年 3 月
提出晶圆级 GPU 的分层分布式页地址转换方案，提升内存系统可扩展性。
导师: Prof. Yifan Sun
论文: HDPAT: Hierarchical Distributed Page Address Translation for Wafer-Scale GPUs
R04
GPU 性能工具链
Scalable Architecture Lab
2024 年 9 月
为 AMD GPU 设计动态二进制插桩工具，研究GPU加速器的性能特征。
导师: Prof. Yifan Sun
论文: Luthier: A Dynamic Binary Instrumentation Framework Targeting AMD GPUs
Education
University of Illinois Urbana-Champaign
Master of Computer Science
Highlights
- GPA: 3.93
- Focused on high-performance computing and GPU optimization.
New York University
B.S in CS, Minor in Math
Highlights
- GPA: 3.79
- HPC Assistant of the NYU High-Performance Computing Team.
Honors
- University Scholar
- Tandon Dean’s List
- Cum laude
教育背景
伊利诺伊大学香槟分校（UIUC）
计算机科学硕士（MCS）
亮点
- GPA：3.93
- 专注高性能计算与 GPU 优化。
纽约大学（NYU）
计算机科学学士，数学辅修
亮点
- GPA：3.79
- NYU 高性能计算团队助理。
荣誉
- University Scholar
- Tandon Dean’s List
- Cum laude（优等荣誉）