NVIDIA's Approach to GPU Training Is Similar to the GDC Presentation

2min Snip

00:00

Play full episode

Summary

Transcript

Episode notes

NVIDIA approach by necessity because their chips and hardware is built at a smaller scale. We train large models by dividing the large models into smaller pieces and gluing it back together with software. They take some of the heavy lifting away from you by giving you libraries to help with this work. And they built a technique called Megatron LM, which has been very influential in distributed GPU training. Microsoft built a system called DeepSpeed. If you add,. if you glue together Megatron, DeepSpeed, Zero and all these techniques and you're really elite, you can get it to work eventually. But it is just so much crud work that no one wants to do.

Scaling laws are as important to artificial intelligence (AI) as the law of gravity is in the world around us. AI is the empirical science of this decade, and Cerebras is a company dedicated to turning state-of-the-art research on large language models (LLMs) into open-source data that can be reproduced by developers across the world. In this episode, James Wang, an ARK alum and product marketing specialist at Cerebras, joins us for a discussion centered around the past and the future of LLM development and why the generative pre-trained transformer (GPT) innovation taking place in this field is like nothing that has ever come before it (and has seemingly limitless possibilities). He also explains the motivation behind Cerebras’ unique approach and the benefits that their architecture and models are providing to developers.

“All great science is reproducible, and AI is the great empirical science of our decade, so we wanted to make sure it’s reproducible.” — @draecomino Key Points From This Episode:

Why Cerebras attracted James.
James explains the concept of wafer-scale computing and why it is so advantageous in the AI space.
A historical overview of large language model (LLM) development.
What James believes to be the most significant natural law that has been discovered in this century.
Why Cerebras wants to get state-of-the-art LLM data into the hands of as many people as possible.
The Cerebras GPT law.
Looking towards the future of LLMs.
Standard practice when it comes to training an LLM (and the problems that developers have been battling for years).
The potential advantage of Cerebras CS2 chips and computers.
Understanding the concept of disaggregated architecture.
How the Cerebras approach differs from the approach taken by other companies (NVIDIA and Dojo, for example).
Cerebras offerings that are available to be used by the public.
The current GPU cloud shortage (and why this adds to the appeal of Cerebras software).
Why the progress being made in the GPT space is incomparable to developments that have come before it.
Potential directions that the world could be heading in as a result of AI developments (and why James is optimistic about it all).
The AI use case that is keeping James up at night.