Matmul optimization is bounded by data movement, not multiplication count
Discussions argue that matrix multiplication performance is dominated by data movement costs rather than multiplication counts, making PRAM-style models outdated. Strassen-like algorithms underperform standard matmul on modern systolic arrays and suffer from numerical instability.
2026-08-19 ~ 2026-08-19 · 4 related posts
- Matmul Optimization: Using Data Movement in Registers as Cost Function — cosminnegruseri · 2026-08-19
- Numerical Instability in "Fancy" Matmul Optimization Methods — francoisfleuret · 2026-08-19
- Matmul Optimization Bottleneck: Data Movement, Not Multiplications — yaroslavvb · 2026-08-19
- Strassen's Algorithm Worse Than Standard Matmul on Systolic Arrays — yaroslavvb · 2026-08-19