Intel enters the laptop discrete GPU market with Xe Max

If you want a <=75 watt GPU, which happens to be the limit for the PCIe slot without an external power connector, you have very limited options. For nVidia, the 1650 is the only modern part that will fit the bill. For AMD, you have to reach back to the RX 500 series, but good luck finding one. They're extremely uncommon even among the lower-end 500 series, as many of them require an external connector anyways.

AMD do have lower-powered "workstation" offerings, but they're so expensive that it's hard to imagine putting one in a workstation voluntarily. They're also extremely poor performers, to the extent that you might not really see much of a benefit over a Vega iGPU.

The Xe Max is only 25W TDP, which is very attractive if you're hoping to build a passively cooled Linux machine with a low-wattage power supply.

Are they though? The GPU embedded in my ryzen 4900 seems to be just as good, if not better than this xe. The desktop version I would expect to show even better.

That is not entirely accurate though.

From a total GPU market presumptive, Low End / Low Power equals integrated GPU. And both AMD and Nvidia simply cant compete with that. The market of a Low End GPU is tiny and unprofitable. When most consumers or customers are happy with their Intel iGPU.

I dont blame them for lack of SKUs in this segment. And AMD is competing with their own APU as well.

People looking for Linux laptops certainly do. I've tried the AMD route twice the last ten years and I've been disappointed by crappy driver support and battery life in both cases.

To justify it existence, a discrete GPU need to be better than that.

I have to guess that's because the low end GPU market is not very profitable.

I did. My laptop has i3-6157u CPU, I chose it due to the Intel Iris GPU there.

If I would be looking for a new one today, for similar reasons I would have picked an AMD APU.

I'd like to counter this. Intel's driver quality is going down for a year or so. My old EliteBook 850G2 is running with Intel Graphics but, it started to show some flickering and corruptions for no reason. Nobody bothered to fix it yet.

Intel's e1000e drivers started to behave awry on some older cards and, Intel didn't fix it. On the contrary, they continued to pile patches over this buggy driver. Kernel guys and Debian developers had to revert some patches back to stabilize the driver again. It still loses connectivity sometimes.

By older cards, I mean some of them are barely two years old.

IF, a big IF, Intel gets its thing together and restores the quality, it's great. Otherwise, I don't put them on a pedestal for driver quality.

On the Windows side, I have an Intel wireless card which is officially supported by current drivers but, the new drivers bork the card after standby and system has to be powered off to restore card connectivity. The computer is a HP Spectre X2 convertible and, only older, out of the box Windows drivers (which are stock Intel drivers) can utilize the card correctly.

I feel like Hip, OneAPI, ISPC, etc. haven't improved anything over CUDA.

With CUDA we had one proprietary API. Now we have 3-4.

The problem is how this intersects with the dynamics with the Intel organisation. Firstly, there's already five teams that already do whatever you want to do, so you go up the chain saying "I need 5 engineers to do X" and the answer comes back down the chain "Team Y already has 20 people, they do X, you should get them to do it". But so now rather than doing X, you're paying Team Y to do X, oh and Team Y doesn't give a shit about X, they care about Z, so about half your budget for X is now being subtly shifted to work on Z. So the cost of your project massively inflates.

Oh, and because when your project was approved they canned something else, some senior executive is saying that half your headcount should be "internal transfers" so suddenly your project has to find work for 50 software engineers at Folsom who just finished working on Intel's Modem team (before that they worked on Intel's previous GPU team too before that got canned).

Now comes the real issues, because you've got all these different teams that you rely on for your project and they don't answer to you, they answer to random different org structures, it's practically impossible to hold them accountable. Which leads to hilarious dynamics where team A will fall behind on their tasks, but they already know team B has fallen behind, so team A will lie about being able to deliver, because they know they'll never have to. So the critical path on your project isn't the team that's fallen behind, it's the team that's fallen behind and all the other teams that have made the calculated judgement your work can be deprioritized because team B will sink your project anyway.

Finally comes the release. All of these internal fights don't change the fact that you're responsible for delivering a product to market, and almost certainly if it's an important project it'll be part of some Group's goals (like the Datacentre group or networking group). Now, the General Manager of the group isn't going to miss her/his goals. So you'll "ship" a product, by which we mean you'll send a handful of sample products to a partner. They probably wont work, they'll be a tiny subset of the functionality you originally promised. So anyway, that's the point at which someone realises we've just spent 10,000 years of engineering effort to deliver a discrete Graphics chip that is marginally less powerful than the integrated one.

But the product is out there, so you've got sales data now and revenue. So you need to agree your roadmap for the future! But the first product was shit, and getting it into any sort of state to be competitive would be an enormous amount of work because it was compromised at every point, so no one will sign off on throwing good money after bad and so what happens? Nervana? Fuck you! We're betting on Habana Labs now!

But Intel has already committed to using active silicon interposer instead of abf for substrate, which is way more expensive.

They are probably attempting this because for a long time OEMs would include both Intel's iGPU (not like they had a choice) but also an Nvidia GPU, even if it had the same or lower performance than the iGPU, just so Intel doesn't fully monopolize their devices.

Then again, by this logic not sure why any OEM would pay twice for Intel GPUs, either.

The problem is neither the old or new method of multi-gpu has ever been really successful at actually working in the app you want it to.

That you can’t do multi-GPU means you can’t combine GPUs to make them into a NUMA GPU cluster that is better performing than any of its nodes.