← Browse

World's Fair 2024

System Design for Next-Gen Frontier Models — Dylan Patel, SemiAnalysis

Dylan Patel

Overview

This talk discusses the system design challenges and future directions for running and training next-generation large language models. It highlights the significant computational and memory bandwidth requirements for inference, particularly for models with trillions of parameters. The presentation also touches upon the massive scale of infrastructure being built for training these models and the associated engineering hurdles.

Who should watch

Key takeaways

Notable quotes

*We haven't seen a big capabilities leap... models that we're using today are largely the same as the models that were trained in 2022.*
*Decode is extremely memory bandwidth intensive... you have to load the whole model from the weights... for decode.*
*Running inference at batch size one is horrendously expensive... it's great to run if you're running it on your own personal devices but if you're running it in the cloud.*

Watch on YouTube →

Unofficial community note. Prefer the recording for nuance.