In the field of high-performance computing, there is a clear divergence between the needs of AI and those of traditional science. Low-precision formats like FP8, useful for training neural networks, introduce unacceptable errors in physical or climate simulations. To close this gap, researcher Laslo Hunhold is developing a specific format that prioritizes numerical accuracy without neglecting the performance and energy efficiency demanded by these calculations.
The Development of a Standard for Exact Scientific Computing ⚙️
Hunhold's work focuses on creating a floating-point format that optimizes bit usage for the range and precision required by scientific models. Unlike AI formats, which sacrifice dynamic precision for speed, this proposal seeks to ensure numerical stability in long iterations. The goal is a design that integrates into specialized hardware, reducing the energy consumption of supercomputers without compromising the reliability of results in critical research.
When Your Climate Simulation Prefers Not to 'Hallucinate' Results 😅
It's understandable. While an AI can generate an image of a cat with six legs and no one is surprised, a nuclear fusion model that invents a new state of matter could cause quite a stir in the lab. Apparently, in science, they prefer their calculations not to be creative, but obstinately exact. So, for now, let's leave FP8 for the dreams of neural networks and use real bits for real problems.