if only they would put the GPU ...


:rockout:RIGHT SIDE UP! :roll:

Seen rumors mid range xe is supposed to run multi GPU with new iris pro.
Would be a killer move,amd should have done it long time ago.
www.notebookcheck.net/New-evidence-of-Intel-s-multi-GPU-support-for-upcoming-Xe-discrete-cards-uncovered-in-Linux-drivers.440252.0.html

i haven't been following intel on releasing the discrete gpu.
Are these consumer workstation gpu? or something else?

cucker tarlsonSeen rumors mid range xe is supposed to run multi GPU with new iris pro.
Would be a killer move,amd should have done it long time ago.
www.notebookcheck.net/New-evidence-of-Intel-s-multi-GPU-support-for-upcoming-Xe-discrete-cards-uncovered-in-Linux-drivers.440252.0.html

ermm why and how?
As the article mentions, this has been claimed before, heck multiple iterations of the idea of multiple different gpu's working together, but it has never actually come to anything good/usable.

Next to that...AMD should have done this a long time ago..with the APU's? why?
The more powerful cpu's from AMD dont have a build in gpu and those that do are semi low end, meant for the office work or laptops etc, there it make sense.
To pair that with a dedicated gpu is a bit silly really.

On the article itself, more non-news from Intel it seems:
Intel Marketing: "We won't let you forget us"

cucker tarlsonSeen rumors mid range xe is supposed to run multi GPU with new iris pro.
Would be a killer move,amd should have done it long time ago.

AMD has done a lot of R&D on this, so has Nvidia and so has Intel. There are fundamental issues with how GPUs work today that prevent from this being effective for gaming (and other real-time rendering applications). Either someone has to come up with a new approach (and collectively they have tried a lot of things) or something has to change about how graphics are rendered/computed - a new paradigm and APIs probably.

Rumors are pretty much always there as there is constant research going on around multi-GPU solutions. Crossfire and SLI were/are the most efficient methods for now. Again, this is for gaming and other real-time rendering applications. GPGPU or similar applications (non-real-time rendering), especially data centers and distributed GPGPU with lower latency/time dependency are much easier and have been largely figured out for a long while (research into improving these has obviously not stopped and is ongoing and this are better in that area as well, the pain points and bottlenecks are simply different).

cucker tarlsonSeen rumors mid range xe is supposed to run multi GPU with new iris pro.
Would be a killer move,amd should have done it long time ago.
www.notebookcheck.net/New-evidence-of-Intel-s-multi-GPU-support-for-upcoming-Xe-discrete-cards-uncovered-in-Linux-drivers.440252.0.html

Only Vulkan and D3D12 can do that but virtually no developers take the time to implement it. As far as I know, only Sniper Elite 4 supports it. And that's hardware agnostic: would work across an NVIDIA GPU, AMD GPU, and Iris Pro integrated if all were in one machine.

ZoneDymoOn the article itself, more non-news from Intel it seems:
Intel Marketing: "We won't let you forget us"

As long as Intel brings more competition to GPU space we should be happy. When they said they are bringing out GPUs we knew that it would take a couple years or more. Some progress reports - especially significant ones like this - are definitely welcome.

FordGT90ConceptOnly Vulkan and D3D12 can do that but virtually no developers take the time to implement it. As far as I know, only Sniper Elite 4 supports it.

A simplified version of the problem is that while Crossfire and SLI (and their DX12 counterpart - implicit mGPU) do multi-GPU largely automatically and developer needs to enable it and test it to make sure they avoid a bunch of gotchas, doing this in Vulkan or DX12 (implicit mGPU) requires developer to basically implement it.

fancuckerHopefully the final nail in AMDs coffin and the first true Nvidia competitor after years of inefficient regressive architectures. RDNA can barely compete against 12nm with a new node and even AMDs ashamed to release a hot 5800XT

ok slow down there alt account, I know it does not matter if this one gets banned because you have your main account and can just make a new alt account, but still, have a bit more respect for yourself.

londisteAs long as Intel brings more competition to GPU space we should be happy. When they said they are bringing out GPUs we knew that it would take a couple years or more. Some progress reports - especially significant ones like this - are definitely welcome.

idk to me this is just non news, if you want an update for every little screw they put into the card then we can have a LOT of articles about this.

We knew it was coming, we knew the expected release date and...nothing has changed except for that we know its coming a bit more certainly now? yay.

Delays would be news but reinforcing what we already know really isnt, unless there were previous rumors that the originally expected date was not going to be met.

It was possible a long time ago to do Crossfire using the GPU in the APU. This isn't new.

londisteA simplified version of the problem is that while Crossfire and SLI (and their DX12 counterpart - implicit mGPU) do multi-GPU largely automatically and developer needs to enable it and test it to make sure they avoid a bunch of gotchas, doing this in Vulkan or DX12 (implicit mGPU) requires developer to basically implement it.

Crossfire/SLI work completely different from what Sniper Elite 4 has. Namely, it sends all of the data to all of the cards and each card focuses on rendering part of the scene then puts together the parts and displays it. It's an extremely inefficient way to do it and severely punishes the best card paired in an asymmetric configuration.

What Vulkan and D3D12 can do is schedule work based on warp load. A Xe card, for example, could get 25 frames scheduled while the Iris Pro gets 5 frames scheduled per second. I'm not entirely sure how that works in regards to VRAM but all of the GPUs work on a need-to-know basis rather than cloning like SLI/Crossfire does. It's much more efficient from a performance an power consumption perspective.

@FordGT90Concept, exactly. The key here is that this work scheduling has to be managed by developer.
I have not looked very closely into what exactly the couple available explicit mGPU games do in terms of memory usage (and I do not have Crossfire or SLI right now) but VRAM is not the biggest concern with it right now.

Yup, and despite that being available for many years now, only Sniper Elite 4 does it as far as I know. So unless Intel has something new up their sleeve, this isn't exciting at all.

VRAM is high bandwidth and high latency. In GPU design, VRAM performance is the number one factor they have to engineer around. Every clock that the VRAM is able to respond, they have to have all of their requests queued while still (hopefully) executing on previous data. Anything that hinders VRAM performance can severely reduce work throughput.

In before devs start making fun of Intel.

Scheduling 25 frames on dGPU and 5 on iGPU is tried-and-true approach and pretty much what Crossfire/SLI do, minus differences in scheduling. Intel actually had the better (or at least more innovative) idea a few years back when they were talking about mGPU stuff - getting iGPU to do postprocessing effects. Basically dGPU does most of the frame and last few postprocessing effects are done on iGPU. These effects need less resources (textures, buffers, basically things in memory). It didn't work all that well but the idea is sound. Managing the data, scheduling and performance levels of different GPUs throughout this is still a pain though.

I think this is getting off the main topic, sorry.

londisteScheduling 25 frames on dGPU and 5 on iGPU is tried-and-true approach and pretty much what Crossfire/SLI do, minus differences in scheduling.

It is not. The dedicated GPU would produce frames at roughly the same rate the iGPU does so you'd get about 90% uplift from iGPU performance. If the dedicated GPU is capable of producing frames faster than 190% of iGPU performance, then you're better off leaving Crossfire/SLI disabled. As I said, Crossfire/SLI split the screen and each card renders its area of the screen then puts them together. It cannot take advantage of asymmetric hardware.


Post processing effects do make some sense.

londisteAMD has done a lot of R&D on this, so has Nvidia and so has Intel. There are fundamental issues with how GPUs work today that prevent from this being effective for gaming (and other real-time rendering applications). Either someone has to come up with a new approach (and collectively they have tried a lot of things) or something has to change about how graphics are rendered/computed - a new paradigm and APIs probably.

Rumors are pretty much always there as there is constant research going on around multi-GPU solutions. Crossfire and SLI were/are the most efficient methods for now. Again, this is for gaming and other real-time rendering applications. GPGPU or similar applications (non-real-time rendering), especially data centers and distributed GPGPU with lower latency/time dependency are much easier and have been largely figured out for a long while (research into improving these has obviously not stopped and is ongoing and this are better in that area as well, the pain points and bottlenecks are simply different).

Uhm, but since both SLI and Crossfire are scaling like 90-99% with dual GPU in several games (so the potential is there), what is the this R&D work focused on? Making the scaling 100% all time on all the games?

FordGT90ConceptIt is not. The dedicated GPU would produce frames at roughly the same rate the iGPU does so you'd get about 90% uplift from iGPU performance. If the dedicated GPU is capable of producing frames faster than 190% of iGPU performance, then you're better off leaving Crossfire/SLI disabled. As I said, Crossfire/SLI split the screen and each card renders its area of the screen then puts them together. It cannot take advantage of asymmetric hardware.

Crossfire and SLI are doing either AFR or SFR. With current rendering technologies, AFR seems to be the more common method of choice. Both work on a level of frame (or two frames). 25 on dGPU and 5 on iGPU example is on a higher level and is effectively same solution but with better scheduling. Scheduling in this case is a more complex problem for asymmetric hardware than it seems at first glance.

voodooFXUhm, but since both SLI and Crossfire are scaling like 90-99% with dual GPU in several games (so the potential is there), what is the this R&D work focused on? Making the scaling 100% all time on all the games?

It works in some games and fails in many others. SLI/Crossfire users can do custom profiles and are tweaking the settings to make games work better but even that does not always help. An mGPU solution needs more reliability in how it works in terms of what games it works in.

Currently there are games that scale perfectly but there is an increasing number of games where it does not scale or scales very poorly. That takes away a lot of the value.

TheLostSwedeHey, I have an idea, why don't you start a GPU company and make something that breasts the lot of them, for a quarter of the price? Everyone here would be eternally grateful to you...

Adding to this - IP and patents are going to be a very serious problem for any new player in the GPU market. It might be worth checking the various lawsuits between companies making GPUs. AMD, Intel and Nvidia are all covered with extensive cross-licensing agreements. Other vendors have been mostly bought up and surviving separate ones are generally in the mobile GPU market - ARM's Mali, Qualcomm's (old AMD-sourced) Adreno and Imagination's PowerVR.

It may not be competition to high-edn like 5700XT or 2070, but just small graphics aimed to use in notebooks.

As for release date, this is first silicon, and may require few revisions before go to shops. Each revision take some 5 months (3m for manufacturing, 2m for analyzing and preparing next revision)

My guess is their top tier dgpu will be rx570 tier..which is not bad for first time..I just want AMD to make a rx580 tier APU (1080p 60fps)..do it and 80% gamer will forget about dgpu.

londisteAdding to this - IP and patents are going to be a very serious problem for any new player in the GPU market.

Actually I am wondering why no one tried licensing ARM Mail GPUs and make something custom with them.

I miss 3dfx. That is all.

whats the improvement over the i740, why havn't intel told us, i want to know how much preformance uplift from their last dedicated card

cucker tarlsonSeen rumors mid range xe is supposed to run multi GPU with new iris pro.
Would be a killer move,amd should have done it long time ago.
www.notebookcheck.net/New-evidence-of-Intel-s-multi-GPU-support-for-upcoming-Xe-discrete-cards-uncovered-in-Linux-drivers.440252.0.html

AMD actually DID it... Results were zero to unimportant (equal to discrete card, or <5% better).

Whole dual-card idea sounds good, but realization depends on devs, and rarely they've invested time in making dual-card games with significant gain. Most of them actually had no gain at all. "The rest" were AAA titles, and probably sponsored by one of manufacturer to include dual-card optimization.

I also don't expect much of first cards, yeah there is Raja-Raja-Raja all over, but if we compare this with AMD... Supposedly, Raja had *some* influence in optimizing old architecture, and his real contributing is showing from now on... Time he spent in Intel is much less, and there are those rumours about "dual-Iris" or "double Iris" that inspire no confidence at all, for me at least.

It may be different for some future generation, but this one - I really, really doubt. Rajas' expertise is nice motivation, but one man doesn't make top products for a long time. Besides, NVIDIA/AMD have, like, 10 zillions of various GPU patents for decades, it's really narrow maneuvering space for Intel. Again, not forever, but 1st generation - definitely.

Sweet, I’m pretty excited for an Intel stand alone card. If not for performance at least because they are an entry.