> Fixing the willful ignorance that the market structure today is very different from 20 years ago
Indeed. And his first action was to diss their own AI efforts. Because AI is just some niche area that they can ignore.
Just as Battlemage GPUs were getting decent reviews and sold above the MSRP.
The craziest thing is...
Intel could totally try to capture the LLM market 'bottom up' if they wanted to.
As an underdog in the GPU market, all they need to do is start by making cheap boards with lots of VRAM. I'm talking 32GB boards under 1k.
They don't have to be fast. They just have to take these bigger models into VRAM and be fast enough that it's better than dealing with normal CPU+RAM.
That gets them into the market, and then they can follow up with more expensive and 'enterprisey' silicon that is faster for the data centers.
Alas, that's probably too alien a thought for Intel, as they prefer that thick margin...
This is such an obvious gap in the market for GPUs right now. Nobody wants to make an affordable card with a ton of vram. 32GB isn't even enough. Someone needs to make 48 or 64gb gpus for reasonable prices. Surely GDDR isn't that expensive. AMD and nVidia margins must be insane.
This doesn't seem like a bad idea but let's follow it a few steps. If a key tactic is to take share by shipping LLM-friendly consumer GPUs, one question is would this work. Setting aside the technical issues, they'd certainly sell some. They'd be limited by software, that's still much better on Nvidia, so it would be people with a near-term need for inference at lower cost than Nvidia's going rate. Two things to think about:
1) How might Nvidia etc respond? They've made one-off SKUs for crypto, they could certainly respond quickly with a part that matched on memory but had much better software (meaning, more compatible with tools and better performance. AMD doesn't have the software, but their hardware is find and they could similarly up on-board memory. So Intel would really have to compete on price.
2) Ok, now we've found some 2nd or 3rd place success in a business built on logic fabbed at TSMC and DRAM from Samsung or Micron. If this is the future, why have fabs or any of the associated R&D?
I don't know what the right answers are but maintaining Intel at anything resembling its current size seems like a pretty tough puzzle.
I like this point. Maybe needs pci4 + large vram + mid line gpu + cheap cpu. The cpu could maybe go scatter/gather/atomics like nics so getting data into and out of gpu is offloaded from the mb cpus doing app work.
What's your source for that? All I've read is that he recognizes that Intel has not had been selling competitive amounts of product in the core areas where Nvidia and others are making most of their money. You could debate why that is but it's certainly not that they've been ignoring AI.