2 September 2026, 14:30-16:00

With the emergence of various AI model scaling laws, a critical question arises: Can existing hardware support the growth of these models, or will there be a significant game-changer that can perform the same tasks at radically lower costs? For these potentially groundbreaking or bold hardware designs, we are interested in estimating their impact on both system-level power and performance, as well as model-level accuracy.

The AIxSIM collective, comprising experts from Imperial College London, the University of Cambridge, and the University of Edinburgh, strives to develop a collection of open-source instruments and frameworks dedicated to overcoming the cross-stack simulation obstacle. This endeavor is financed by the Advanced Research and Invention Agency (ARIA) as a part of the Scaling Compute program.

This tutorial consists of the following talks:

Introduction to AIxSim
Binglei Lou, Imperial College London

AI4HW and HW4AI: The PLENA Ecosystem Journey
George Haoran Wu, University of Cambridge

Specialized Agents for Close-to-Metal Kernel Generation on AI Hardware Benchmarks, Harnesses, and Self-Improvement
Jiayi Nie, University of Cambridge

SysSIM: AI Cluster Performance Simulation
Luo Mai, University of Edinburgh