Systems & HPC

Multi-GPU ANN Search

Extended the CAGRA graph ANN algorithm into a multi-GPU vector-search framework: sharded scatter-gather, zero-copy shared-memory loading of SIFT1B, 50M+ vectors indexed at >97% Recall@10.

PythonCUDAcuVS/CAGRACuPySLURM

View the code

8.9× build speedup

Flight data

From the cluster runs

Index build time vs. dataset size, 1 vs. 2 GPUs (SIFT)

This one needs NVIDIA GPUs on the ECE Nebula SLURM cluster, so it can’t run here. The charts are drawn from the project’s own benchmark_results.csv — the raw output of those cluster runs.