Dynamic Graphs on the GPU

on YouTube ↗58m

" summary

Claims a practical GPU stack: GPUs have 300,000 threads in flight, ~144 SMs, ~80 GB HBM and ~36 MB L2, and the PCIe/CPU link (~32 GB/s) often bottlenecks real pipelines. The group implements cuckoo hashing on GPU and per-vertex hash tables instead of CSR rewrites, shows a 1000x slowdown from contention when naively porting CPU b-tree latching, and reports hash-table and bulk-rebuild strategies that work well for large batched updates.

AI-written summary · about

" chapters (19)

1. 0:00Introduction and context
2. 0:31Session welcome and speaker
3. 1:05Bridging theory and systems
4. 3:49Systems research perspective
5. 6:51GPU programming model
6. 12:41Hardware architecture
7. 16:12GPU memory system
8. 21:51Latency tolerance
9. 25:00Performance challenges
10. 28:20Memory coalescing
11. 30:23GPU and CPU interaction
12. 33:10The role of machine learning
13. 35:58Graph processing on GPUs
14. 38:48Graph work engineering
15. 42:15Dynamic data structures
16. 44:55Data structure design
17. 48:00Hash tables and evaluation
18. 52:30Future research directions
19. 53:49Q&A and conclusion