Linus Torvalds: Discrete GPUs are going away

[Insert_Nickname]

So lets say you can double the EUs of the Haswell GT3 to 80-90EUs for Broadwell GT3 and keep the iGPU die size the same at 100-110mm^2. Do you actually believe a 80-90 EU Broadwell GT3 (Gen 8) even with 128mb eDRAM will be able to match the HD7870 or GTX660 ??
Broadwell GT3e will not even reach GT-650Ti levels of performance. GT-750ti will still be way faster and cheaper.

For my own part, I'll believe the performance of these 80/96EU GT3/4e's when I see it. Come on, even the top-of-the-line GT3e/Iris Pro 5200 hasn't even caught up with a lowly GT640 in performance. Next generations top-of-the-line may hit GTX650 (perhaps even 650TI) levels. Only problem is that dGPUs have moved ahead to GTX750(TI) level performance. Those cards use the lowest-end GK107/GM107 chips in Nvidia's line-up. I do think that says everything needed...

There is an additional thing few seem to have caught on. One of the reasons Intel's IGPs have made such dramatic progress recently is simply because they until recently have been extremely poor performance. Its much easier to improve a poor architecture then it is to improve something that is already pretty efficient (Maxwell/GCN). Like as not performance improvements will flat line matching AMD/Nvidia's architectures at some point.

 

[witeken]

thats your assumption, simpler circuits, higher densities , different trade offs in "storage" vs compute etc. complex devices are never defined by one metric. Have you every tried to undervolting a GPU? all mine have pretty big head room downwards, likely about maximizing yields, so even changes in binning could have a big impact.

Lower voltage will give you higher efficiency but lower clock speeds (performance), so you need more cores to make up for that, just like Apple has done with Haswell's GT3 graphics in the Macbook Air.

Nvidia's GK110 already has a die size of 550mm², the wafer costs will go up at 20nm and 16FF and yields will be worse.

So the voltage won't drop to levels where chips benefit most from FinFETs.

 

[witeken]

There is an additional thing few seem to have caught on. One of the reasons Intel's IGPs have made such dramatic progress recently is simply because they until recently have been extremely poor performance. Its much easier to improve a poor architecture then it is to improve something that is already pretty efficient (Maxwell/GCN). Like as not performance improvements will flat line matching AMD/Nvidia's architectures at some point.

Exactly, that's my points. It's also what ARM's Jem Davies said in Anand's hangout. The big difference however, is that the TDP delta between a console and smartphone is more than 1 order of magnitude, while the difference in TDP between a desktop CPU and low-end/midrange GPU isn't big at all.

 

[Qwertilot]

The other thing to remember is that Intel's fastest iGPU to date hasn't really been trying in terms of desktop power consumption.

It looks hard to tell just how much power the iGPU can pull down but it doesn't look too huge. A 65W TDP chip at most though, so if they pushed it out to ~100W and gave all of that power budget to more iGPU you'd be much closer.

Clearly no interest in doing so as yet though, we'll see if they're developing one.

 

[Racan]

If VR manages to take off, how will Intel and it's low watt integrated graphics parts manage to power this revolution?

https://www.youtube.com/watch?feature=player_detailpage&v=dxbh-TM5yNc#t=1286

 

[ShintaiDK]

If VR manages to take off, how will Intel and it's low watt integrated graphics parts manage to power this revolution?

VR actually looks to lower the requirements. No 4K etc to deal with. Most of them target 1080p as the future for VR.

 

[witeken]

If VR manages to take off, how will Intel and it's low watt integrated graphics parts manage to power this revolution?

https://www.youtube.com/watch?feature=player_detailpage&v=dxbh-TM5yNc#t=1286

Thousands of pixels per inch and hundreds of megapixels? This guy is seriously overestimating the requirements for VR.

 

[f1sherman]

If VR manages to take off, how will Intel and it's low watt integrated graphics parts manage to power this revolution?

https://www.youtube.com/watch?feature=player_detailpage&v=dxbh-TM5yNc#t=1286

VR actually looks to lower the requirements. No 4K etc to deal with. Most of them target 1080p as the future for VR.

you should click that link :sneaky:

 

[Alatar]

VR actually looks to lower the requirements. No 4K etc to deal with. Most of them target 1080p as the future for VR.

100% incorrect. It's a well known fact that VR is one of the technologies that benefits most from the increased resolution. Oculus for example knows perfectly well that even 1080p isn't quite enough. They've made many comments in the past about perhaps getting a higher than 1080p resolution on the first consumer release. And definitely going higher than that in the future.

60fps is also the absolute minimum for VR. Even the DK2 that's going to get shipped soon is 1080p 75hz. First consumer version is most likely higher than that.

And on top of that you're going to have to also drive much higher FoV than you normally would on a PC monitor.

Intel will be lucky if they can match a lowly 750Ti in the next 5 years. Right now their $600 chips aren't even capable of matching a 640. Let alone driving actual high end games or other GPU bound applications.

 

[witeken]

Intel will be lucky if they can match a lowly 750Ti in the next 5 years.

Skylake GT4, which will be released within 12 months, will easily have 2X as much FLOPS as a GTX 750 Ti.

 

[Insert_Nickname]

Skylake GT4, which will be released within 12 months, will easily have 2X as much FLOPS as a GTX 750 Ti.

Its not just a question of raw theoretical performance, but also how efficiently that performance is used in practice. Practice is where Intel usually fall flat on their face. That's not even mentioning how poor their graphics drivers are.

F.x. I did a little test swapping my own HD7870 (2560 GFOLPS) with a burrowed GTX750TI (OC'ed to 1345MHz/1721.6 GFOLPS). Know what? The 750TI had ~90% of the peformance, despite it having only ~67% the raw GFLOPS performance of the 7870. Its not like it was CPU limited either on an OC'ed 3770non-K@4.3GHz. Theoretical FLOPS does not tell the whole story...

It also goes a long way to show just how efficient Nvidia's architecture really is. I almost can't wait for the "real" Maxwell's to arrive. One of those is going in my rig on launch day, anyone else who wants to play games with their piddly IGP are welcome to do that.

 

[witeken]

Haswell GT3e has 20% more FLOPS than GT 650M, but GT 650M performs on average about 1.3x better or so (because Iris Pro is 1.5x behind at high quality). So even when we take the difference in flops->fps efficiency into account, that's still not enough to be worse than GTX 750 Ti. And I'm comparing Maxwell to Gen7.5. Gen9 will obviously be a lot better.

 

[AtenRa]

Broadwell won't have 80-90EUs. It will have 48EUs, so the total die area of the APU with the uncore, 8MB of cache and 4 cores will be well within 130mm² (GTX 750 Ti alone is already 148mm²).

And here comes the ugle thruth of the 14nm higher cost. Intel cannot double the transistor count in the first year of the new process due to higher cost. They will only introduce Broadwell GT3 with 48EUs (Haswell has 40EUs) and GT2 with 24EUs (Haswell has 20EUs).

Broadwell GT3e with 48EUs will maybe reach GK107 (GT-650) performance level. But GT650 was released in 2012, Broadwell GT3 will be released in 2015. So why people believe iGPUs are closing the gap ?? they are still 2-3 years behind.

 

[NTMBK]

VR actually looks to lower the requirements. No 4K etc to deal with. Most of them target 1080p as the future for VR.

The second Oculus Rift dev kit is already at 1080p, and rumours have the final unit at 1440p.

 

[Insert_Nickname]

Haswell GT3e has 20% more FLOPS than GT 650M, but GT 650M performs on average about 1.3x better or so (because Iris Pro is 1.5x behind at high quality). So even when we take the difference in flops->fps efficiency into account, that's still not enough to be worse than GTX 750 Ti. And I'm comparing Maxwell to Gen7.5. Gen9 will obviously be a lot better.

You're talking about a GT650M? In that case there are 3 possible configurations, the highest one of which is only good for 691.2 GFLOPS. Even then as you write, the GT650M has on average 1.3x the performance of Iris Pro. The desktop GTX650 has a 17% performance advantage (Same GK107@1058MHz=812.5 GFLOPS) over that. The desktop GTX750TI has an 88% (1306 GFLOPS) performance advantage over the GTX650M. While Intel may catch up with the GT650, I don't think they can pull an 88% increase in performance in one generation.

Plus you overlook the fact that we're talking about Intel catching up to a three year old discrete GPU at Skylakes launch. Hardly what I'd call impressive...

 

[witeken]

There will be a GT4, which will have 2.4X as much cores as Haswell GT3.

Nvidia improved their already cutting edge architecture by 1.35X per core and 2X per watt, according to their own slide, so Intel could certainly do something similar or better with 2 iterations.

 

[dangerman1337]

The second Oculus Rift dev kit is already at 1080p, and rumours have the final unit at 1440p.

Yeah, Oculus VR (or Palmer at least IIRC) has said they want CV1 a higher res and are thinking future versions to go even higher than 4K.

 

[sm625]

Of course discrete GPUs are going away. They've been going away for 15 years, and 15 years from now they'll... you guessed it... they'll be going away. The problem with IGPs is that they are junk, for whatever reason they are just junk. No IGP can even come close to competing with a HD7750, and that is but a mere tiny gpu that could easily fit onto a cpu. What's the problem? RAM? Then put RAM on the package. OK but that costs an arm and a leg, and its money spent for no purpose. Why does it need to be so expensive? Why is it $30 for 2GB of GDDR5 on a video card but then some outrageously inflated value to put that same amount of RAM onto a cpu package? And more importantly, who cares? Discrete is cheaper and better, for whatever reason, and intel charging $200 extra dollars for their joke of a solution isnt going to change that. Intel charging massive amounts of money is probably the sole reason why IGPs havent taken over.

 

[NTMBK]

No IGP can even come close to competing with a HD7750,

Kaveri performs roughly the same as a DDR3 7750. With fast RAM and a bit of overclocking it can even outperform it, as a DDR3 7750 only has 1600MHz memory.

 

[sm625]

Kaveri performs roughly the same as a DDR3 7750. With fast RAM and a bit of overclocking it can even outperform it, as a DDR3 7750 only has 1600MHz memory.

DDR3 on discrete is kind of irrelevant. The HD7750 GDDR5 smokes Kaveri. Double the passmark score. Double the framerate in games like bioshock infinite. It will be 2-3 more years before any IGP under $200 will outperform a 7750. There might be some $500 iris super mega pro that does it next year for $666, but I'm not counting that.

The main point I'm trying to make is that when we move beyond GDDR5 to on package memory using TSVs or whatnot, it is still going to be cheaper to bring this technology up along the path of the discrete video card, because there is competition there, rather than intel greed which will only stifle technological progress.

 

[Throckmorton]

What's the market share of Linux? In the real world I mean. People who do work on computers sure don't use Linux

 

[f1sherman]

What's the market share of Linux? In the real world I mean. People who do work on computers sure don't use Linux

they do... students, scientific community, academia. just to mention some

but yeah, too much tinkering can be unhealthy

 

[itsmydamnation]

So the voltage won't drop to levels where chips benefit most from FinFETs.

And AMD has a 438mm sq chip that uses more power, you are ignoring what im saying just to repeat the same simplistic statement. If going to 0.8 volts gives the best overall performance then thats where they will go, how much wattage/power they consume (your original statement), has nothing to do with it.

VR actually looks to lower the requirements. No 4K etc to deal with. Most of them target 1080p as the future for VR.

yes ~90 fps @ like 1440P is easy to hit.......

 

[antihelten]

If VR manages to take off, how will Intel and it's low watt integrated graphics parts manage to power this revolution?

https://www.youtube.com/watch?feature=player_detailpage&v=dxbh-TM5yNc#t=1286

foveated rendering

 

[Fox5]

What's the market share of Linux? In the real world I mean. People who do work on computers sure don't use Linux

People using Linux for work was the primary use case for Linux prior to Android taking over the mobile phone market.

So the market share of Linux? Probably more than Windows at this point if you include Android.

More than Windows server in the server world.
And pretty heavy in the scientific and engineering fields.