Vai al contenuto
AI.info

events

Faster AI: GPU Rooflines, Agent Learning & Homa’s Fast Transport Protocol

Meetup online sulle prestazioni dell’IA, con interventi sull’utilizzo delle GPU e sulle curve roofline di vLLM, sull’apprendimento dalle tracce degli agenti e sulla latenza di Homa per i messaggi di…

Faster AI: GPU Rooflines, Agent Learning & Homa’s Fast Transport Protocol | Eventi sull’IA

In inglese

Starts
2026-10-19 16:00 UTC
Organizer
AI Performance Engineering Meetup (Palo Alto)
Location
Online event

The AI Performance Engineering Meetup (Palo Alto) hosts an online session with three technical talks. Rajat Dandekar uses visual explanation and code to show how concurrent users affect GPU utilization, linking vLLM measurements on an H100 to the roofline model. Rustem Feyzkhanov discusses turning agent traces into reproducible simulations, verifiable rewards, and model improvement through fine-tuning or reinforcement learning, including when to change prompts, tools or harness instead. John Ousterhout explains Homa’s approach to small-message latency and its open, workload-dependent relevance to AI. Organizers give introductions and updates.

Event website