Journal
Recent Posts
Notes on GPU programming, local AI, and distributed systems.
clusterllm
Building a Heterogeneous Inference Cluster over USB4
Splitting a model between Intel and AMD GPUs: what crosses the cable, where latency accumulates, and what a successful two-node run actually proves.
strix-halollm
LLM Inference on Strix Halo
What a 64 GiB Strix Halo system can tell us about local inference: memory budgets, speculative decoding, context costs, and the limits of a benchmark.
cudahip
Setting Up HIP with CUDA
Some notes on setting up CUDA and HIP for a system with an NVIDIA GPU or with an AMD GPU.
hpcgoptimization
HPCG: 3. Sparse Matrix Formats and Memory Access Optimization
Understanding sparse matrix formats with simple python code. By using ELLPACK format for representing sparse matrices, we can optimize the memory access pattern of the HPCG benchmark.