Speaker
Description
The AI boom is reshaping processor design. Hardware vendors now prioritize high throughput for matrix multiplications and dense linear algebra, optimize aggressively for low precision, and integrate specialized units such as tensor cores. Meanwhile, compute performance continues to grow much faster than memory bandwidth, and latency improvements lag behind both. The result is a widening gap between computation and data movement — a shift with profound implications for scientific simulation software.
Traditional simulation codes rely heavily on IEEE double precision and were designed for architectures with more balanced compute and memory characteristics. In an AI-driven hardware landscape, these assumptions are increasingly challenged. Should we embrace mixed precision? Is double precision becoming a luxury? Do we need emulation for higher precision, or decouple memory and arithmetic precision through compression? Should we trade communication for recomputation to reduce data movement?
This talk explores how hardware trends driven by AI affect simulation software design and argues that we must rethink — though not abandon — our numerical and architectural assumptions to remain efficient on emerging platforms.