I actually did email Deepak Patil (head of Intel Graphics division) about this around a year ago, haha. Never did get a response though.
It is something that is easy to miss if you are just looking at typical business strategy and finances. A high memory consumer GPU would undercut their server GPUs, which are presumably higher margin intended golden geese. It's easy to see them chasing server markets and "gamers" being an afterthought.
However there is huge demand right now for a modern, even a crappy modern, GPU with gobs of memory. Make the card and the open source AI tooling for it will start sprouting in days after it's release.
It's an extremely powerful position to have every at-home AI geek's setup to be bound to using intel cards and intel focused tooling. Nvidia and AMD won't do it because they want to protect their server cards.
> It's an extremely powerful position to have every at-home AI geek's setup to be bound to using intel cards
So, incredibly small market share while your competitors already have the first-mover advantage and nailed down the ecosystem? With no data backing it up, I think, graphics cards for local LLM needs is not really on demand. Even for gaming it’s probably more attractive, but then again, that’s not even where the real money is.
>So, incredibly small market share while your competitors already have the first-mover advantage and nailed down the ecosystem?
Exactly. This x100. It was easy for Nvidia to succed in the LLM market by winging it, in the days when there was no LLM market, so they had the greenfield and first mover advantages.
But today, when Nvidia dominates the mature LLM market, Intel winging it the same way Nvidia did, won't provide nearly the same success as Nvidia had.
Ferruccio Lamborghini also built a successful sports car company by building tractors and cars in his garage. Today you won't be able to create a Lamborghini competitor with something you can build in your garage. The market has changed unrecognizably in the mean time.
The market share is incredibly small but also incredibly well aimed.
The people learning how to do local LLMs will be the people directing build out of on-prem transformers for small-midsize companies. The size of the market is irrelevant here, it's who is in that market and the power they will have that is extremely relevant.
> ..open source AI tooling for it will start sprouting...
AMD has tried this for many of its technologies and I don't think it is working. Granted, they suck at open sourcing, but a shitload of it was open sourced. See TinyGrad voyage into the Red Box driver (streams on youtube).
Intel doesn't have to open source anything. People will build everything needed to run intel cards efficiently as there is currently zero options for affordable video cards with high memory.
It's either old slow Tesla cards with 48GB or $2000 nvidia cards with 24GB.
> People will build everything needed to run intel cards efficiently as there is currently zero options for affordable video cards with high memory.
I think you're overestimating what people can and will do.
Nvidia didn't succeed because it just launchend cards and let people write CUDA for them. Nvidia is where it is because it has an army of researchers and SW engineers developing the full stack from research papers, to frameworks, to proofs of concepts, showing customers the value of paying for their pricey HW + SW, most of it proprietary, not community developed.
"People" alone won't be able to get even 10% there. And that's ignoring the fact that Nvidia HW is not FOSS so they'd be working blind.
It is something that is easy to miss if you are just looking at typical business strategy and finances. A high memory consumer GPU would undercut their server GPUs, which are presumably higher margin intended golden geese. It's easy to see them chasing server markets and "gamers" being an afterthought.
However there is huge demand right now for a modern, even a crappy modern, GPU with gobs of memory. Make the card and the open source AI tooling for it will start sprouting in days after it's release.
It's an extremely powerful position to have every at-home AI geek's setup to be bound to using intel cards and intel focused tooling. Nvidia and AMD won't do it because they want to protect their server cards.