Vai al contenuto
AI.info

events

Make Your GPU Earn Its Keep: A Kernel Engineering Deep Dive

Rajat Dandekar di Vizuara spiega come la profilazione, il modello roofline e la gerarchia della memoria delle GPU guidino lo sviluppo di kernel per l’IA più veloci, dal tiling a FlashAttention.

Make Your GPU Earn Its Keep: A Kernel Engineering Deep Dive | Eventi sull’IA

In inglese

Starts
2026-11-30 17:00 UTC
Organizer
AI Performance Engineering Meetup (Munich2)
Location
Online event

The AI Performance Engineering Meetup hosts an hour-long online session on how GPU kernels get faster. Following welcome updates from Chris Fregly and Antje Barth, Rajat Dandekar of Vizuara walks through the roofline model, CUDA execution and the GPU memory hierarchy, then shows how profiling guides optimisations such as coalesced memory access, tiling and tensor cores, with attention and FlashAttention as examples. The aim is a practical method for deciding what to measure and whether a change helped, followed by audience questions. No prior CUDA or C++ experience is required.

Event website