How machinesactually work.
I’m most interested in what happens beneath the software: memory, data movement, and parallel hardware, especially as AI pushes all three to their limits.
Developing a Cost-Effective and Scalable Contextual AI
With G. Ecklund and Dr. M. Holliday, I helped build a multi-model retrieval-augmented generation system for academic advising. I designed its storage layer on PostgreSQL, ChromaDB, and Redis, and deployed it in Docker on Apple Silicon with Metal.
We tested it on 100 WCU advising questions, ten runs each, with every one of the 1,000 answers graded by subject-matter experts. The paper was peer reviewed, and we presented the poster at CCSC South-East.
Memory is the bottleneck.
Modern AI is limited less by how fast a chip can compute than by how fast it can move data to where the computing happens. My research interest is that gap: memory hierarchy design, data locality, parallelism, and computing closer to memory.
The chart shows why. Each step away from the processor costs more time. Switch to human time to feel the difference.
| Level | Real time | If 1 nanosecond were 1 second |
|---|---|---|
| Register | 0.3 nsabout a third of a second | |
| L1 cache | 1 nsabout 1 second | |
| L2 cache | 4 nsabout 4 seconds | |
| Shared cache | 12 nsabout 12 seconds | |
| Main memory | 100 nsabout 2 minutes | |
| NVMe SSD | 20 µsabout 5½ hours |
Approximate figures on a log scale. Real numbers vary by chip.
Parallel computing
Splitting one job across many cores, a GPU, and a cluster. Hover over or tap a card to watch the work spread out.
Eigenvalue solver
Rust, threads sharing state through Arc and Mutex
28% faster than the instructor’s reference, measured in GFLOP/s
Matrix multiplication
CUDA on an NVIDIA V100
Thousands of GPU threads working on one product
SUMMA
MPI across nodes on Clemson’s Palmetto cluster
Distributed-memory matrix multiplication
Mandelbrot
Pthreads and ISPC
Threads and SIMD lanes splitting the image
A Simplified Approach to Accessible PDFs Using LaTeX
A talk at Western Carolina University on how clemson.sty makes accessible documents the default for anyone writing in LaTeX.