Linus Torvalds: Discrete GPUs are going away

[ultimatebob]

Sadly, the 3D graphics drivers are so bad in some Linux distributions for some newer cards that you're actually better off using the integrated graphics.

 

[_Rick_]

Uhm, memory bandwidth? Or are you assuming HBM / stacked DRAM already?

That's the one problem, that I kept addressing in my previous posts.
But still, it's only an economic issue.
High bandwidth memory interfaces aren't voodoo, on cards they see frequent use, and they could easily ported to integrated designs.
They are economically problematic, at least for the foreseeable future, since DDR4 will be too slow for streaming applications, but GDDR5 and successors will be too expensive. Quad channel memory controllers are a step in the right direction, but provide nightmare logistics, and wider buses hit tracing limits.
On cars with 12GB of memory maximum, this can be somewhat mitigated, but in a shared memory architecture, memory volume also needs to scale up.

A temporary solution would be to use a large high bandwidth volatile cache (8-16 GB @ .5-1TB/s) L4/5 cache. Making the memory hierarchy deeper adds overhead, but with only an ~x10 differential in speed at each step from L1 down to HDDs (L1-L2-L3-on-die RAM - off-die cache VRAM - RAM - PCIe cache SSD - storage SSD - HDD) there are more options for economic scaling.

 

[BrightCandle]

As far as I know DDR4 is basically going to bring about a peak of 100 GB/s. So to get into discrete territory the iGPU is going to need either quite a large cache or its going to need a lot of stacked memory for the iGPU part of the chip. Either will do the job but both are going to increase the cost of the CPU quite a lot.

 

[BFG10K]

Oh, wait, you're saying "in graphics performance". Sorry, I thought your graph compared CPU FLOPS vs GPU FLOPS; I misunderstood the graph.

The point is that underlying process advantages mean nothing without architectural information to back it up, along with real-world performance. You keep talking about theory and things that don’t exist yet as if they’re reality.

Intel claims a 75X gaming improvement since 2006, outpacing both Moore's Law and your graph:

The difference between a 2006 part (7100GS) and today's flagship GTX780Ti is a lot more than 75x. For example, the texturing fillrate alone is 140x times higher (1.4 GTexels vs 196). When the 7100GS runs out of VRAM the reduced performance would be hundreds of times slower, for example in 4K performance.

To say nothing of multi-GPU performance which is not only much faster but has multi-slot physical/electrical budgets which a single consumer CPU socket will never have.

They're not falling behind further, and certainly not exponentially. And if you want to compare FLOPS, I think even a theoretical Gen7 IGP with 72-144 EUs can give you a decent understanding of how Intel will catch up in the coming 1-2 years.

“I think”? “I think” is not evidence of anything. Please show us gaming benchmarks of these non-existent parts you keep referring to as fact.

I already told you. Intel will improve its microarchitecture so that it isn't much behind anymore, and because its manufacturing lead is expanding, Intel will be able to get much better IGPs than what would have been possible without this 2-3 node advantage. Just look at how good or bad GPUs were 2-3 nodes (or 4-6 years) ago.

Again, manufacturing doesn't mean anything without details of the underlying hardware and how it actually performs in the real world. We were told amazing things about Larrabee and look how that turned out.

Why do you have to refer to the situation multiple years ago? Why is that relevant? Roadmaps change, plans change, targets change, all sort of things change. If you're going to refer to the past, when Intel was much more behind, you're always going to come to the conclusion that Intel will never catch up, obviously.

You keep talking about things that don’t exist while I keep talking about things that have already happened.

A GTX Titan? I think 10nm is very likely: 10nm is 5x more dense than 28nm, so your massive 550mm² GTX Titan is reduced to 110mm². Add the CPU and you APU is about the size of Ivy Bridge/Haswell. I don't know how high it will be able to clock, but note that Intel will use germanium at 10nm, which could potentially quite dramatically reduce improve consumption and performance.

“Very likely”.
“Potentially”.
“I think”.

Again, show us facts, not your opinion and/or musings.

If/when any of this happens, dGPUs would've advanced a magnitude (again, based on past history which has actually happened) so that a Titan will be completely obsolete, just like the 7100GS is now.

 

[witeken]

The point is that underlying process advantages mean nothing without architectural information to back it up, along with real-world performance. You keep talking about theory and things that don’t exist yet as if they’re reality.

Better process = higher density -> more transistors available + lower price; better performance; lower power consumption.

If Intel is at that highest point, Nvidia will be 2-3 points away in the coming four years or so. It won't be easy to make up for that disadvantage. It's guaranteed that this will hurt Nvidia. At this moment, Intel is serious about IGPs, but their products simply don't show that, yet. Iris Pro is not available in any LGA SKU, and they haven't updated their 2 year old architecture.

The unlocked GT3 Broadwell SKU with a much improved Gen8 architecture and 20% more cores/GT will be the first sign of that. Not much later, there will be a Skylake part with GT4 Gen9 graphics.

“I think”? “I think” is not evidence of anything. Please show us gaming benchmarks of these non-existent parts you keep referring to as fact.

You keep talking about things that don’t exist while I keep talking about things that have already happened.

“Very likely”.
“Potentially”.
“I think”.

Again, show us facts, not your opinion and/or musings.

What I say are facts: there really will be Gen8 and Gen9 graphics for Broadwell and Skylake. Broadwell and, depending on how you interpret the rumor, Skylake will have more EUs/GT. There will be a GT3 and GT4 SKU with much more EUs than the 20 we have now.

How you and I interpret this information, however, is opinion and apparently we disagree. So I can't show you benchmarks.

$If/when any of this happens, dGPUs would've advanced a magnitude (again, based on past history which has actually happened) so that a Titan will be completely obsolete, just like the 7100GS is now.

This is what we do different. You extrapolate the past, I use the information I have about the Gen8/9 SKUs and manufacturing advances.

And I my opinion, I think my approach is better. GPUs can't advance nearly as much without better process nodes, and those will become more rare in the future while current information suggests Intel still isn't slowing down.

So yes, GPUs will still become better. There will be a node shrink to 20nm and a change to FinFETs shortly thereafter, which will benefit GPUs. However, after that, there won't be another die shrink (without die shrink you can't put more cores in a GPU, which means no increase in performance) until late 2018 or later. Current information even suggests that although there will be a much higher density in 20nm, the price per transistor will hardly decrease at all, which is the same as no shrink at all.
In the same time, Intel will also shrink their transistors to, to the 14nm node and later to the 10nm and even to the 7nm node, before 2019. That are ~3 die shrinks more. Meanwhile, Intel will also improve transistor performance and power with new transistor materials, which will make up for TSMC's catch-up FinFETs.

This will correspond to a lead of 4 or more years. If very difficult to be competitive when you're that much behind. A lot of Nvidia's die area and TDP (which is exclusively available for GPUs) in the future will be used to catch up to Intel's high-end IGPs.

We'll also have to see how this translates into end-user prices.

 

[sontin]

nVidia has been saying that they need to change the architecture to improve the perf/watt because smaller process nodes are not bringing the same effect as it was in the past.

Maxwell v1 is doing it: 2x over the same Kepler GPU without a new node.

nVidia doesnt need the latest process technology to easily beat Intel.

 

[witeken]

And who says Intel didn't have a big team working on Gen8/9/10 and AMD on GCN2.0? Nvidia isn't the only company that is able to make an efficient GPU. If everything about the GPU is the same (or equivalent), then the best process node wins.

 

[sontin]

Why would Intel invest the same amount of money into their GPU tech when there isn't a payoff? iGPU is a free goodie and nothing you can monetize.

 

[naukkis]

Why would Intel invest the same amount of money into their GPU tech when there isn't a payoff? iGPU is a free goodie and nothing you can monetize.

BUt there is, CPU-wise intel hasn't bring any noticeable improvement after Sandy Bridge, GPU-wise they have. Intel's most valuable comsumer chips today are those with Iris Pro graphics....

 

[witeken]

I don't know Intel's motivations. A CPU with Iris Pro will obviously be more expensive (not free goodie). Maybe they want consumers to buy their CPUs with good IGPs instead of GPUs. I think mobile is probably important for them. They can recycle the IGP in phones, tablets, laptops and desktops, so it has value over a good range of products and for desktops they can simply add EUs (Gen7 is modular with slices).

Edit, from Iris Pro review:

Intel hired some very passionate graphics engineers, who always petitioned Intel management to give them more die area to work with, but the answer always came back no. Intel was a pure blooded CPU company, and the GPU industry wasn’t interesting enough at the time. Intel’s GPU leadership needed another approach.
[...]
Pure economics and an unwillingness to invest in older fabs made the GPU a first class citizen in Intel silicon terms, but Intel management still didn’t have the motivation to dedicate more die area to the GPU. That encouragement would come externally, from Apple.

It came from Apple, but with Brian Krzanich as CEO, I think he sees more value in IGPs than Otellini: “If it computes, it does it best with Intel.”

 

[sontin]

BUt there is, CPU-wise intel hasn't bring any noticeable improvement after Sandy Bridge, GPU-wise they have. Intel's most valuable comsumer chips today are those with Iris Pro graphics....

And yet only a few products exist which have Iris Pro...

I don't know Intel's motivations. A CPU with Iris Pro will obviously be more expensive (not free goodie). Maybe they want consumers to buy their CPUs with good IGPs instead of GPUs. I think mobile is probably important for them. They can recycle the IGP in phones, tablets, laptops and desktops, so it has value over a good range of products and for desktops they can simply add EUs (Gen7 is modular with slices).

iGPU is a free goodie for consumer. Nobody is going out and pay more for a notebook with an iGPU over the same model with a dGPU.
And the fastest iGPUs come with the bigges die and they cost much more than a CPU with a much slower iGPU and a dGPU.

Edit, from Iris Pro review:

It came from Apple, but with Brian Krzanich as CEO, I think he sees more value in IGPs than Otellini: “If it computes, it does it best with Intel.”

And Apple is one of the few companies which are using a CPU with Iris Pro...

 

[Qwertilot]

Of course they can monotise iGPU. They'll sell desktop options against an equivalent CPU/dGPU combination. Gives them quite a bit of leeway price wise.

Might be plausibly attractive with Broadwell K and onwards, we'll have to see
(Performance of course but also smaller, simpler, quieter computers etc.).

Apple of course huge leverage with a fair chunk of the high end computer market and the ability to move away from Intel entirely if they get really annoyed with them.

The thing which worries me a bit about maxwell is that, if they can ~double power efficiency just from tweaking the architecture then what on earth were they doing when they designed the previous one?!

Well that is/was obviously very competitive too, so maybe GPU architectures are just incredibly hard or something.

 

[ShintaiDK]

Iris Pro was prohibited from the LGA socket with 22nm CPUs due to the combined physical size. Both Broadwell and Skylake adds Iris Pro to the LGA socket.

Both will also increase the amount of Iris Pro SKUs substantially. And it fits the classic history of integration.

 

[jpiniero]

Why would Intel invest the same amount of money into their GPU tech when there isn't a payoff? iGPU is a free goodie and nothing you can monetize.

Well, if they are going to lockout and/or discourage dGPU usage, they have to make the iGPU better; otherwise OEMs will revolt.

The purpose of course is to kill off nVidia.

 

[ShintaiDK]

Its not about killing nVidia. Its simply about performance/watt, package size and integration. You simply cant make as small laptops as we see if you had to use dGPU. Or as small NUCs for example. nVidia is just a casualty of legacy.

 

[VulgarDisplay]

It will definitely be interesting the next few years to see what happens with GPU market share. I could foresee nvidia losing a large amount once all their dgpu's in laptops become irrelevant.

 

[AtenRa]

Heh, Intel is turning to become a GPU manufacturer at a fast pace. Soon the majority of its ICs (in the consumer market) will devote more space for the GPU than CPU.

 

[njdevilsfan87]

Eh, I'd argue all day that architecture trumps process any day.

See: NVIDIA Tesla v.s. Xeon Phi (Knights Corner)

With the Phi, does this means if that dCPU is also going away? It seems like dGPU is being merged with dCPU are being merged. Actually, I wouldn't even call it that. People are just realizing GPUs are pretty damn good as functioning as the dedicated device of the entire system. I believe Xeon Phi can run without a CPU, and I think Nvidia's Maxewll will be the first capable of doing so as well.

I think in the future we will still be able to custom build desktops. However, we won't be buying a CPU and GPU separately. We'll be buying all in on devices. The cloest equivalent today would be buying a Celeron + crappy iGP or a Xeon Phi to run everything on your system.

 

[Fox5]

As far as I know DDR4 is basically going to bring about a peak of 100 GB/s. So to get into discrete territory the iGPU is going to need either quite a large cache or its going to need a lot of stacked memory for the iGPU part of the chip. Either will do the job but both are going to increase the cost of the CPU quite a lot.

If a mainstream IGP comes out with 100GB/s to system memory, game over to most of the discrete market. That'll be in the performance range of anything with a 128-bit memory bus, which is the $150 and under market.

 

[zink77]

Memory bandwidth and bus widths will keep Discrete GPU's around forever. I just don't see discrete going away because everyone who poo poo's discretes forget heat and space. Linus's comment is idiotic because many people were expecting graphics to be integrated into CPU's in the pentium 2 celeron and pentium-3 era which never happened. Guys like linus were saying the same thing way back in the days about 'how soon everything will be integrated', it took damn near forever to get video on CPU's. People were predicted the end of the videocard back in the late 90's and it didn't happen. The naysayers predictions of integration were extremely off, many predicted in the late 90's 2000's everything would be integrated but it didn't happen until almost a decade or more later and even then there are serious trade offs with integration.

It took practically a decade for any video integration into cpu's that had any kind of decent performance and they are still out competed by the cheap low end discrete cards.

Add in boards allow you to do cool things to increase memory bandwidth that you simply can't do on integrated circuits. When you integrate you're trading one set of constraints for another. Given that GPU's have problem domains that aren't going away anytime soon besides gaming discrete won't be going away.

 

[Zodiark1593]

With the Phi, does this means if that dCPU is also going away? It seems like dGPU is being merged with dCPU are being merged. Actually, I wouldn't even call it that. People are just realizing GPUs are pretty damn good as functioning as the dedicated device of the entire system. I believe Xeon Phi can run without a CPU, and I think Nvidia's Maxewll will be the first capable of doing so as well.

I think in the future we will still be able to custom build desktops. However, we won't be buying a CPU and GPU separately. We'll be buying all in on devices. The cloest equivalent today would be buying a Celeron + crappy iGP or a Xeon Phi to run everything on your system.

There really isn't much a modern video card is missing to be able to run solo itself. They're basically self-contained computers in and of themselves designed for highly parallel processing tasks, quite a far cry from rasterizers of old, and even the original Geforce (the first gpu apparently), and are capable of more than just graphics processing. So in a way, Linus was right.

 

[Arkadrel]

Memory bandwidth and bus widths will keep Discrete GPU's around forever.

Only if you dont believe in the "good enough" for x task, stuff people claim.

Like how cpus are fast enough for any daily task for the mainstream user, so building big 200watt cpus isnt needed, thus intel went perf/watt instead.

at some point, unless eye-candy in games accelerates at a unheard of pace, igpu will reach a "good enough" point.

A jump in memory technology like the hybrid memory cube, could really change things.

Once we start seeing HMC take off, prices will go down, and become common place on motherboards (this technology will likely kill off GDDR/DDR4 ect).

And we ll start to see motherboards with 15-20x as much memory bandwidth on them, enough to feed the iGPUs.

Once you have 150-200 GB/s memory bandwidth on a motherboard for the CPU, the iGPU for the "mainstream" user will be fast enough to not need a discrete gpu. (IF intel & AMD make a beefy iGPU)

The only people who will keep useing discrete gpus will be guys that need beyound normal graphics horse power, and with the shrinking market, the prices will sky rocket.
Basically discrete gpus will end up being for guys that build servers like render farms ect.

Lets not forget theres a reason other than just better Performance/watt, Price, for moveing the iGPU onto the CPU.

With HSA or a intel made technology like it at some point, you can see massive improvements in computational performance when doing GPGPU workloads. It also becomes a million times easier to code for.

IT MAKES SENSE TO HAVE THE GPU ON THE SAME CHIP AND INTEGRATED INTO THE CPU.

Theres massive amounts of performance for GPGPU workloads to be found there.
A discrete gpu wont be able to compete in terms of gpgpu workloads soon.

 

[Ayah]

for the "consumer", easily. for enthusiests, eventually but definitely far in the future. for current ~GTX750Ti range people, relatively easily done, but motherboards will shoot up in cost, pin count will rocket up and air cooling is out of the question as this is FX-9590 range power consumption. and obviously, you would have needed several years headstart on this little project.



for people who actually use their computers for computing: I'm pretty sure I'll be dead before I see a high end GPU integrated into a high end CPU: with current gen tech, that's something like a 550W compound chip and most likely over 14 billion transistors.
with current tech, you'd still need to offload heavy compute to dedicated GPGPU or a cluster. the most amusing thing about the compound high end pro chip idea at this time is that you'd most likely need/want somewhere around 450-500GB/s of memory bandwidth. signal integrity will be a nightmare with that pin density.

and to get that memory bandwidth, I'd *love* to see you route that on anything less than 20 layers, 14-18 might be possible but I really doubt it.

as always, even if energy efficiency increases, the high end products simply use that efficiency to add more performance while keeping similar-ish power consumption.
additionally, some problems are still far easier to solve (or in some cases, impossible in parallel) sequentially instead of in parallel.

 

[zink77]

Only if you dont believe in the "good enough" for x task, stuff people claim.

Like how cpus are fast enough for any daily task for the mainstream user, so building big 200watt cpus isnt needed, thus intel went perf/watt instead.

at some point, unless eye-candy in games accelerates at a unheard of pace, igpu will reach a "good enough" point.

A jump in memory technology like the hybrid memory cube, could really change things.

Once we start seeing HMC take off, prices will go down, and become common place on motherboards (this technology will likely kill off GDDR/DDR4 ect).

And we ll start to see motherboards with 15-20x as much memory bandwidth on them, enough to feed the iGPUs.

Once you have 150-200 GB/s memory bandwidth on a motherboard for the CPU, the iGPU for the "mainstream" user will be fast enough to not need a discrete gpu. (IF intel & AMD make a beefy iGPU)

The only people who will keep useing discrete gpus will be guys that need beyound normal graphics horse power, and with the shrinking market, the prices will sky rocket.
Basically discrete gpus will end up being for guys that build servers like render farms ect.

Lets not forget theres a reason other than just better Performance/watt, Price, for moveing the iGPU onto the CPU.

With HSA or a intel made technology like it at some point, you can see massive improvements in computational performance when doing GPGPU workloads. It also becomes a million times easier to code for.

IT MAKES SENSE TO HAVE THE GPU ON THE SAME CHIP AND INTEGRATED INTO THE CPU.

Theres massive amounts of performance for GPGPU workloads to be found there.
A discrete gpu wont be able to compete in terms of gpgpu workloads soon.

All of these are based on hypotheticals, and the biggest roadblock I see is any advance that allows full integration also allows for full integration on add in boards, you keep forgetting that. What applies to one also applies to the other.

Predict that any IGP powerful enough to get rid of discretes will be done long after linus is dead because it requires advances in material sciences that are difficult. The same way flying cars and 'fusion' was just right around the corner. Lets not forget people who thought that Mhz would scale much bigger than it has, P4 was originally slated to scale to 10Ghz then it was found out that wouldn't be the case because of unexpected roadblocks. It's been 10 years roughly since the pentium 4 and we've only managed to squeeze another 1.5Ghz over many processor generations.

I see unexpected roadblocks to IGP displacing discreets and any IGP advance will also apply to add in boards negating the advantage.

 

[ShintaiDK]

I see unexpected roadblocks to IGP displacing discreets and any IGP advance will also apply to add in boards negating the advantage.

Not entirely. They are already decoupled. The IGP simply advances at a faster pace with many more lower hanging fruits to pick.

New GPU uarch or improved uarch every 1-1½ years. A continually expanding node lead. 14nm vs 28nm. (We may even see 10nm vs 28nm)

Stacked memory will only have relatively minor effect on dGPUs. While it will frog leap IGPs.

And again, sooner or later the economic incentive to develop the dGPU further stops due to shrinking volume. Remember, the IGP doesnt have to beat the dGPU. Just destroy its ROI to win.