← Back to context

Comment by JonChesterfield

14 days ago

The GPU scheduler is an arm chip. As in taped out, shipping in volume, has been for years. I think there was a project decades ago that put an arm front end on an x86 core. If industry decides they like the aarch64 isa more than x64 I doubt it would cause much trouble to the silicon team.

> If industry decides they like the aarch64 isa more than x64 I doubt it would cause much trouble to the silicon team.

the problem isn't ISA so much as AMD not having a moat around its CPU revenue anymore. Yes, AMD will have ARM chips... and so will Qualcomm, and Mediatek/NVIDIA, and Marvell. AWS and Facebook and Google have already largely dumped x86 chips entirely. But they haven't dumped them for ARM chips from AMD, or Intel, or Qualcomm, they've dumped them for ones they develop themselves at rates that would be commercially unviable for AMD to compete with as a semicustom product (and those companies would not want that in the first place).

It's not about Intel vs AMD and which of the two is faster anymore. It’s about the fact that intel plus AMD is going to shrink to be a minority player in the market as ARM commoditizes a market from which Intel plus AMD extracted economic rents for decades. That free money train is coming to a stop. Absolutely there is going to be a market for AMD as a consultency business helping others integrate chips (just like semicustom today) but even that is going to be something that many others do etc, and the IP cores themselves certainly don't have to be licensed from AMD anymore, there will be many viable offerings and AMD's margins will be thinner when they provide less of the stack and design services etc. Absolutely there will be people willing to pay for the cores too, but it's going to be a lot less than the people who want something that's Good Enough to just hook up their custom accelerator thing to a CPU (just like AWS Graviton and Google TPU etc).

And while consoles are still going to have legacy support as a major inertia factor, even that is not forever, the ftc leak showed Microsoft bid the next-gen console out as a “ML super resolution upscaling” and RTGI focused ARM thing, right? Especially with how far AMD has fallen behind in software and accelerator integration etc. ARM will eventually break that moat too - especially with nvidia pushing hard on that with switch 2 as well etc. Even these ancillary services that AMD provides are largely important because of the most that x86 provides around anyone else having easy interoperation.

Margins are going to come down, and overall market will grow but AMD will command less of it, and also large customers are going to depart that market entirely and do it themselves (which AMD will also get a chunk of, but a much smaller pie etc than selling the whole CPU etc).

Again, like, the problem isn't that AMD doesn't have a Snapdragon X Elite. It isn't even about whether they're faster than Snapdragon X Elite. It's the fact that Snapdragon X Elite is going to do to client revenue what Graviton and TPU are already doing to your datacenter revenue. It's what ARM and Qualcomm and NVIDIA are going to do to your x86 and graphics SIP licensing revenue, and your semicustom integration revenue etc. Even if they are flatly not as good, they don't have to be, in order for them to collapse your x86 rents and your margin on the resulting design+integration services and products. We will come to realize we didn’t need to pay as much for these things as we did, that integration isn’t as expensive when you have five similar blocks from 5 different companies and three places that can integrate the SOC. That inefficiency is all part of the x86 rents/x86 tax too.

This isn't to say you're doooomed but like, Intel and AMD are not winners from the powerplay that Microsoft is making against x86 right now. You literally are going to be facing 3-4 more well-funded large competitors who can now offer services that are now substitutable for yours in a way that nobody appreciated would happen a year ago. Windows on ARM being a viable proposition in terms of support and emulation changes the game on the x86 rents in client/gaming markets, and companies are moving in to take advantage of that. Intel and AMD have literally nowhere to go but down on that one - they will not control the resulting market anymore, actually they already are losing control of large parts of the datacenter market due to cloud ARM offerings from hyperscale partners. Now it's client too, and foreseeable risk to consoles and semicustom etc.

All of that ancillary revenue was just gated by x86 in the final measure. Now that x86 is no longer a necessity, we’re going to see a significant increase in market efficiency, and parties that are dependent on the largesse of those rents are going to have problems. You can’t rely on graphics being a loss-leader (or low-margin-leader) to sell x86 SOCs, for example.

  • It'll be sad if the x64 lines wind down entirely, or end up on IBM style life support, in end of an era sense. Also fatal to Intel.

    AMD would be fine though. They're currently in first or second place across x64, gpu, fpga. There's dedicated networking stuff and the ai accelerator/ npu things as well. Semi-custom has been a big thing for ages. That's a clean sweep across all forms of magic sand.

    They're also the most open source & collaboration themed of all the hardware players. I think Broadcom building on AMD's infinity fabric is a really big deal; that's collaboration at the hardware IP level.

    Right now everything is obsessed with generative AI and Nvidia building artificial intellectual factories, but if that proves to be marketing nonsense it'll kill Nvidia and AMD will shake it off.

    The moat isn't the x64 patents. It's diversification and execution competence.