News Intel Xe HPG Gaming GPU: Taking on AMD and Nvidia

[JarredWaltonGPU]

https://www.tomshardware.com/news/nvidia-amd-ati-graphics,6311.html#:~:text=AMD, Nvidia Conspire to Price Fixing; Sued By,price fixing of GPU processors and graphics cards.#:~:text=AMD, Nvidia Conspire to Price Fixing; Sued By,price fixing of GPU processors and graphics cards.

I would have to look at where this went but it was enough to get a case started.

Ironically, in the emails they also stated Intel would never release a dGPU.

Like a lot of lawsuits, it probably went nowhere. Or a slap on the wrist for the involved companies. There are a lot of ways to argue it wasn't price fixing, though, and proof is much harder to come by.

No but any performance numbers from Tiger Lake should give us an idea. Not an exact idea but enough to extrapolate and increase based on moving to GDDR6X. And that should be early September for that.

Early Tiger Lake leaks suggest GPU performance roughly on par with AMD's Renoir (Vega 8). Which is basically what you'd expect from the theoretical GFLOPS figures. Tiger Lake is going to be TDP constrained on the 15W parts for sure, but I'm interested in seeing what the 28W (35W?) parts can do.

However with the rumors of the RTX 3090 I think Intel wont quite be in the realm of taking names.

Best-case is if Intel's Xe HPG doubles the EU count of Xe HP -- either by just being bigger (more EUs + RT, drop FP64 and other data center stuff), or if EMIB can effectively make for a dual-GPU solution that behaves as a single GPU. 1024 EU Xe HPG would be in the same 21 TFLOPS ballpark as I anticipate for RTX 3090. Drivers, though ... I'll trust Intel's drivers to deliver on all fronts when I see it happen. They're way better now than 2-3 years ago, but as one example, Horizon Zero Dawn fails to run right now. (A fix is being worked on.)

 

[jimmysmitty]

Like a lot of lawsuits, it probably went nowhere. Or a slap on the wrist for the involved companies. There are a lot of ways to argue it wasn't price fixing, though, and proof is much harder to come by.

Early Tiger Lake leaks suggest GPU performance roughly on par with AMD's Renoir (Vega 8). Which is basically what you'd expect from the theoretical GFLOPS figures. Tiger Lake is going to be TDP constrained on the 15W parts for sure, but I'm interested in seeing what the 28W (35W?) parts can do.

Best-case is if Intel's Xe HPG doubles the EU count of Xe HP -- either by just being bigger (more EUs + RT, drop FP64 and other data center stuff), or if EMIB can effectively make for a dual-GPU solution that behaves as a single GPU. 1024 EU Xe HPG would be in the same 21 TFLOPS ballpark as I anticipate for RTX 3090. Drivers, though ... I'll trust Intel's drivers to deliver on all fronts when I see it happen. They're way better now than 2-3 years ago, but as one example, Horizon Zero Dawn fails to run right now. (A fix is being worked on.)

The lawsuit probably did evaporate but the emails are pretty interesting to read.

I also would like to see higher wattage performance numbers. I think TL should be decent for mobile competition.

I also agree with the drivers. Its strange though. When it comes to chipset drivers they do extremely well. But since they are currently pushing for a dGPU I would expect they put a lot of work into making their driver package a lot better especially with how large their software team is.

 

[TerryLaze]

I also agree with the drivers. Its strange though. When it comes to chipset drivers they do extremely well. But since they are currently pushing for a dGPU I would expect they put a lot of work into making their driver package a lot better especially with how large their software team is.

Wouldn't they have to release a completely new driver stack for the Xe anyway?
It would make sense to not see any improvement for current drivers if they are focused on the new ones.

 

[JarredWaltonGPU]

Wouldn't they have to release a completely new driver stack for the Xe anyway?
It would make sense to not see any improvement for current drivers if they are focused on the new ones.

The overall Xe architecture appears to be quite similar to Gen11 Ice Lake -- just with support for 50% more EUs in Tiger Lake / Xe LP. There are some changes I'm sure, but perhaps not enough to require a fully new stack? But Xe HPG will need a lot more work, as it will have to include ray tracing support and a bigger focus on DX12 / Vulkan. Funny enough, Intel highlighted a "completely new, from the ground up" DirectX 11 driver in the Architecture Day, so that's something as well.

 

[vincero]

All of the above points are good reasons to be cautious, but the biggest detractor for Intel graphics (for me personally) is their consistent and proven history of how well (or badly) they provide ongoing support for the products a few years down the line.

This is something that Intel have been bad at on many fronts, be it upgrades from Windows versions, support for fringe / niche products, etc. - I mean look at the flakey support for the PowerVR powered GMA parts back in the day, and now the Kaby Lake G series - supposedly support for 5 years but that already seems to have gone out the window.

Even some of their own IP used to get maybe a couple of years of 'useful' support and that's it. Westmere Core series processors were still being manufactured / released into 2011 but Windows 10 (released 2015) doesn't support them properly - I can't help but think that was deliberate to essentially orphan off the related GMA X4500HD parts from the Core2 series, and only support newer HDxxxx named parts with newer WDDM type drivers.

Compared to their competition, AMD and Nvidia still provided drivers for products older than that, even if the Win10 drivers were just point releases of the existing Win8 ones just to get the job done and they lacked proper WDDM 1.3/2.0 support.

Right now you can get the latest drivers from nvidia which still support back to Kepler (2012)... 8 years... Intel can only dream of that.

 

[TerryLaze]

Right now you can get the latest drivers from nvidia which still support back to Kepler (2012)... 8 years... Intel can only dream of that.

The latest intel drivers still support haswell and that came out in 2013...all that fuzz for one year of difference?!?!
https://downloadcenter.intel.com/do...phics-Driver-for-Windows-15-40-?product=80939
Came out yesterday.
Version: 15.40.46.5144 (Latest) Date: 8/18/2020

 

[vincero]

The latest intel drivers still support haswell and that came out in 2013...all that fuzz for one year of difference?!?!
https://downloadcenter.intel.com/do...phics-Driver-for-Windows-15-40-?product=80939
Came out yesterday.
Version: 15.40.46.5144 (Latest) Date: 8/18/2020

Yeah, and I wonder if that is partly because of the HD Graphics Pxxxx versions that exist since Haswell forcing better driver support for workstation as per some support agreements, or because the Surface Pro's from MS forces Intel to keep releasing updates for X years or MS are paying Intel to maintain drivers better.
Hopefully this will carry forward with the Xe line-up as the same GPU architecture is used in the CPU line, but I'd still be wary - history is not on their side, and the very recent Kaby Lake G support issues do not help shake that impression.

 

[TerryLaze]

and the very recent Kaby Lake G support issues do not help shake that impression.

Well it's hard for intel to supply support for hardware they didn't make themselves, they would have to reverse engineer AMDs iGPU technology to provide driver updates or support of any kind.
I mean what exactly do you expect intel to do? Drag AMD to court over this?!

 

[vincero]

Well it's hard for intel to supply support for hardware they didn't make themselves, they would have to reverse engineer AMDs iGPU technology to provide driver updates or support of any kind.
I mean what exactly do you expect intel to do? Drag AMD to court over this?!

Well, without knowing specifically what was in the contract between them it's hard to say - the initial impression is Intel were responsible for providing drivers (even if these were repacks of AMD's), probably doing some of their own extra validation, etc. - it seems they lost interest in doing this OR AMD may have changed something in the driver packages which made this harder, but I suspect its the first option.
Now the AMD driver packs do not work, BUT contractually Intel should have the right to then get the drivers from AMD and still repack them as needed if AMD are blocking their own installers from working - its not as if they even need to do it as fequently. But Intel aren't doing this.

If the contract actually stipulates that AMD will allow the driver software installer to work with the Kaby Lake G parts and that is not working and isn't fixed, then yes, surely this would be a slam-dunk court case for Intel. Even though the value is small, it would cost AMD more either way. If Intel do nothing it could cost them instead if users demand a rebate due to lack of support, or actually file a class-action against Intel.

Anyway, we've gone way off topic here. Everybody has been saying "hmmm, Intel driver quality is a bit suspect", which is only partially true - it's hugely better than it used to be in the i8xx/i9xx/Gxxx/Qxxx days - you'll most likely find most apps/games work without glaring visual artifacts, except for the rare few. Even AMD / Nvidia hit a few 0-day issues now and then with new things not working properly / at all.
My concern is if you'll ever get a driver say 3-4 years after Xe mk1 releases for when MS release a new Windows build / driver model requirement before you'll be allowed to upgrade.
My gut would say yes, but history says don't be so sure about that.

 

[JarredWaltonGPU]

Anyway, we've gone way off topic here. Everybody has been saying "hmmm, Intel driver quality is a bit suspect", which is only partially true - it's hugely better than it used to be in the i8xx/i9xx/Gxxx/Qxxx days - you'll most likely find most apps/games work without glaring visual artifacts, except for the rare few. Even AMD / Nvidia hit a few 0-day issues now and then with new things not working properly / at all.
My concern is if you'll ever get a driver say 3-4 years after Xe mk1 releases for when MS release a new Windows build / driver model requirement before you'll be allowed to upgrade.
My gut would say yes, but history says don't be so sure about that.

The big issue with a lot of Intel's previous generation graphics solutions are that they were woefully inadequate in terms of features and hardware. Everything prior to about HD 4000 series (Haswell) was fairly limited in certain Direct3D features -- geometry throughput was a big one IIRC. Even Haswell didn't ever get DirectX 12 or Vulkan drivers, which precludes even trying to run a variety of games. That's probably for the best, since the HD 4600 was and is anemic. It's fine for 1080p office work; anything more and you run into problems. Games? Outside of very light stuff like CSGO or LoL, don't bother. Interestingly, Broadwell (Iris Pro) seems to have even worse drivers than Haswell in terms of being updated to work properly.

Skylake and later GPUs at least brought Intel up to a minimum feature level support that mostly allows their hardware to work with modern graphics applications and games. You still get bugs, or games that fail to run at all (Horizon Zero Dawn), but for the past five years Intel has been getting a lot better about drivers and such. Gen11 and now Xe Graphics intend to carry that momentum forward, and Raja Koduri obviously knows how important drivers are when it comes to GPUs.

Realistically, I think Intel could draw a line in the sand and provide a similar ~8 year window for driver support starting at Xe. The minimum level of performance will hopefully be high enough and the future architectures will hopefully be similar enough that Intel can keep them updated. Edge GPU products like PowerVR and Kaby Lake G were just a nightmare in drivers waiting to happen, and sadly it did.