Panel: What are we doing?

on YouTube ↗19m

" summary

Discussion highlights a persistent gap between research and deployable systems, claiming many sparse advances never reach real workloads and arguing that concrete wins need hardware and compiler co-design. Examples: FlashAttention tiling delivers most of the attention speedup by avoiding reloads; exploiting algebraic rewrites couples algorithm and schedule; vendors add features like fp4 clipping on A100 and single-format accelerators that help some kernels but leave others slow. Pushes for roofline limits, sparse-matrix fusion, and data-aware memory representations.

AI-written summary · about

" chapters (11)

1. 0:00Panel introduction
2. 0:43Sparse computing challenges
3. 3:13Algebraic transformations in compilers
4. 4:35Dense to sparse performance
5. 6:01Hardware for sparsity
6. 7:37Hardware design bounds
7. 8:48Sparse compiler robustness
8. 11:20Hardware and compilers convergence
9. 13:30Compiler distrust in hardware
10. 15:38Hardware-software interface
11. 17:25Performance limits summary