Michael Acton
๐ค SpeakerAppearances Over Time
Podcast Appearances
So this is really about Intel's failure about four years ago to launch a competitive product with the graphics processing units that NVIDIA has risen to power on.
So a graphics processing unit is a chip needed to train and run AI models.
And it really sort of eclipsed the traditional central processing units or CPUs that Intel has dominated in in the past.
And this is Intel positioning itself to launch its own GPU unit.
And there are certain aspects of this chip that make it interesting.
One is the memory.
So Intel is using a different type of memory from the extremely expensive high bandwidth memory that goes into chips from the rivals like NVIDIA and AMD.
It's also bypassing the need for liquid cooling infrastructure.
So if you're building a data center, in theory, you wouldn't need to build this huge network of water pipes and things to make it work.
And then there's this sort of looming idea that Intel might actually build this chip finally in-house in its own factories in the U.S., which also brings down the costs.
It also hasn't ruled out the idea that some version of this chip could sell in China.
And that's a really interesting dynamic given the tensions between Washington and Beijing, whether Intel could succeed where NVIDIA and AMD have really struggled in offering some sort of price-effective chip in China.
So what we know is, and Intel's candid about this, that they're not trying to build a GPU that goes toe-to-toe with NVIDIA and AMD on training the frontier models that we're familiar with, so the chat GPTs of the world, the Claudes of the world.
They're really looking to target the infrastructure for running the models.
So it's early days, but Intel's looking to sort of build a platform of its own.
When we're talking about platform, a lot of people just think about chips.
But really, it's not just about the chip.
It's about building a server that chip goes into.
It's about all of the wires and connecting technology that goes into those data centers that allows the chips to communicate with each other.