Samuel Reeder

Associate Engineer at AMD · ML Systems & GPU Computing

Hi, I’m a fourth-year student at the University of Toronto, studying computer science, mathematics, and statistics. At AMD, I work on the software stack underlying modern AI: GPU kernels, their delivery to users, and the benchmarking infrastructure to measure their performance at scale. On my own time, I enjoy training and running models locally. I’m also into skiing, Formula 1, and sim racing.

Graduating April 2027. Seeking full-time roles.

Resume LinkedIn GitHub

Work

AMD / hipDNN & MIOpen

5 days→3 hours

Exhaustive performance sweep duration

Built a Python and Kubernetes framework on Azure Kubernetes Service to distribute model sweeps across 256 GPUs during my AMD internship.

Reported workflow result: roughly 40× shorter elapsed time for the full sweep, not an individual GPU kernel.

Benchmarking infrastructuren shapes → k per subset → ⌈n / k⌉ pods · 256-GPU AKS cluster

Associate Engineer

May 2026 – present

Lead development of hipDNN’s Python benchmarking infrastructure, including time-series tracking and nightly parallel sweeps over 1,197 kernel-shape cases across three GPU architectures. Authored an RFC and implemented runtime pass-by-value tensors, enabling AITER FlashAttention prefill and paged-attention decode kernels with up to 1.8× speedup. Implemented a file-driven kernel descriptor loader, replacing hard-coded engine tables with validated runtime descriptor discovery.

Explore hipDNN

AGS Libraries Software Development Intern

May 2025 – May 2026

Added six hipDNN graph operators spanning MoE, convolution data-gradient and weight-gradient, RMSNorm, and block-scale quantization, with attribute and edge validation. Integrated Convolution operations and BatchNorm fusions in the MIOpen engine alongside reference implementations, and extended MIOpen’s Convolution Fusion C API with additional tensor layouts, delivering a 6.5% speedup across aggregated convolution workloads.

240+ PRs landed · 350+ reviewed · Four AMD Director’s Spotlight Awards

Projects

AsciiNet

Python / PyTorch / CUDA / ConvNeXt / Transformers

Designed and trained a 28M-parameter ConvNeXt/transformer from scratch to turn images into ASCII art in one pass. Distilled a subject-aware glyph-selection pipeline into the network, then paired it with SANA-Sprint to turn text prompts into pictures and pictures into characters.

24 ms / image · 0.31 GiB peak VRAM

At 80 columns on an RTX 5080: 4.5× faster than the teacher pipeline (108 ms/image). Image conversion only; prompt generation and model loading are separate.

Evaluation and training loss

At 80 columns, AsciiNet answered 157 of 480 held-out, five-choice recognition questions correctly (32.7%); its teacher answered 170 (35.4%), a 2.7-percentage-point gap. The judge was Qwen3-VL-8B; chance is 20%. These are automated recognition scores. The examples below are curated illustrations.

Final training loss: approximately 1.503, averaged over the last 100 steps of the second training phase, including 1.043 glyph cross-entropy. This measures how closely the model fits its teacher's targets; the recognition test above compares the legibility of their outputs.

Source Model weights

SANA-Sprint generated illustration of a butterfly. — Saved model outputs, generated locally. SANA-Sprint produced the source illustrations; AsciiNet chose every character. Switch examples and widths to compare the same image at two resolutions.

SANA-Sprint generated illustration of a helicopter. — Saved model outputs, generated locally. SANA-Sprint produced the source illustrations; AsciiNet chose every character. Switch examples and widths to compare the same image at two resolutions.

Generative Piano AI

Python / PyTorch / MIDI / Docker

Curated a MIDI dataset of 100+ classical pieces, designed and trained a transformer model, and fine-tuned it with reinforcement learning on human ratings. The model generates note sequences, durations, and velocities.

Source

Waveform of the generated piano sample. — Recorded model output. The original audio renderer cuts a note’s reverb when the next begins; this sample is a demonstration, not a quality benchmark.An instrumental piano sequence generated by the model. There is no speech. The optional video shows terminal windows running training and generation commands; it is a workflow recording, not a notation or keyboard animation.

Caspr

TypeScript / Next.js / Three.js / Firebase

Built a causal-graph visualization platform for University of Toronto AI researchers, leading a seven-developer student team. Created for the U of T Machine Learning Group.

Source Recorded team demo

Caspr application: a 3D causal graph with labeled connected nodes, a details panel, and search and filtering controls. — Authentic application capture from the team repository. JSON graph data becomes an interactive 3D view with search, filtering, and sharing.

MIPS Tetris

MIPS Assembly / Bitmap graphics

Implemented a complete Tetris game in 1,200+ lines of 32-bit MIPS assembly, including native 2D graphics, piece rotation, line clearing, score tracking, and a next-piece queue.

Source Run it yourself

Original gameplay recording, converted from GIF to a controllable silent video. Run the assembly source in Saturn or MARS.A silent, 16-second recording of tetrominoes moving, rotating, and stacking in the MIPS bitmap display.

Background

University of Toronto · HBSc

Computer Science Specialist in Artificial Intelligence, with Statistics and Mathematics minors. Expected April 2027.

AMD

May 2026 – present

Associate Engineer

Lead hipDNN benchmarking and API work. Built one-command agentic harness setup tooling adopted as standard onboarding by 9+ teams, and presented two talks to roughly 100 AGS Libraries engineers.

AMD

May 2025 – May 2026

AGS Libraries Software Development Intern

Landed 240+ PRs and reviewed 350+ more across ROCm’s public C++ GPU-library repositories, spanning graphs, GPU backends, validation, logging, and distributed performance testing.

DataAnnotation Tech

May – Oct. 2024

Data Analyst & Programmer

Authored and debugged RLHF code corpora in C++, Rust, and Python. Analyzed prompt-response datasets for distribution shifts and label noise.

GoChain

May – Oct. 2022

Software Engineer Intern

Redesigned the GoTrace frontend, debugged the Go-based GoChain blockchain node daemon, and developed gochain.io for outreach and recruiting.