Pl/cuda the fusion of HPC grade power
summary
Adds procedural C-blocks inside PostgreSQL CREATE FUNCTION that compile to GPU kernels and run on ZPU/CUDA devices, with infrastructure to allocate device buffers and pass SQL args/results. Introduces AR/Matrix columnar arrays to pack columns for efficient 32‑byte GPU memory transactions and functions to convert between rows and 2D arrays. Demonstrates k‑NN Jaccard similarity over a 10M‑compound catalog and claims a 10050× speedup versus a CPU C version; also mentions SSD→ZPU DMA as future work.
AI-written summary · about
chapters (7)
- 0:00PL/CUDA introduction
- 7:43Why PL/CUDA was designed
- 12:51PL/CUDA overview
- 25:44Similarity search in drug discovery
- 39:27K-means clustering for traffic
- 46:31PL/CUDA features and challenges
- 49:46Q&A
related talks
18mThe "New Normal" or An Evolution
35mAccelerating PostgreSQL Full Table Scan OLAP queries
46mKeynote: When Hardware and Databases Collide
50mFrom Transactions to Intelligence: Postgres for Open AI & Analyti
28mCAPI/OpenCAPI Accelerated PostgreSQL
52mNUMA vs PostgreSQL
30mParticle physics, 10,000 times faster
32mHacking Postgres Executor For Performance