How will GPUs perform post-14 nm?

[Mondozei]

So if I am using Nvidia's roadmap as an example, this year we'll get 20 nm Maxwell unless something drastic happens. Then we'll get refresh next year and 14 nm Volta in 2016. Considering that 10 nm and further are at this point no longer on Moore's law, which we are already seeing on CPUs, how will GPUs perform beyond 2016?

Can we still count on ~30% performance increases per year as we've been spoiled for several years now or are we going to see ~10% increases like we've seen for CPUs?

 

[KaRLiToS]

How can we know when we don't even know how 20nm will perform or overclock?

 

[Ed1]

So if I am using Nvidia's roadmap as an example, this year we'll get 20 nm Maxwell unless something drastic happens. Then we'll get refresh next year and 14 nm Volta in 2016. Considering that 10 nm and further are at this point no longer on Moore's law, which we are already seeing on CPUs, how will GPUs perform beyond 2016?

Can we still count on ~30% performance increases per year as we've been spoiled for several years now or are we going to see ~10% increases like we've seen for CPUs?

First I would not count on 16nm by 2016 , If we are lucky your only getting 20nm by 2015 area and each time going smaller it gets harder, costly .

Since GPU is mostly parallel processing the core count can go up with similar die size and slightly better voltages (hopefully) . So for sure once things get to point they can't go smaller it is going to be harder to increase core count w/o increasing die to a point of thermal limits and power limit .

I think its way to early to try and speculate as to performance .

 

[ShintaiDK]

We are still to see 20nm GPUs. Then we get 20nm with finfets called 16nm later on. Then after that we may see a real 16nm. But that may be in 2018 or so.

In terms of scaling. GPUs scale fine with a process shrink due to the way they work. However as the process nodes from the foundries are more and more low power focused. It may be more performance/watt related, rather than pure performance. So wait and see.

Moores law work fine for CPUs. Remember, Moores law is not about performance.

 

[NTMBK]

Hopefully they will all benefit nicely from unified address space, what with being integrated into the CPU.

 

[wand3r3r]

Precisely between 0 and 5000% better. I'm sure by now you see there are no accurate ways of doing anything more than guessing.

 

[R0H1T]

Precisely between 0 and 5000% better. I'm sure by now you see there are no accurate ways of doing anything more than guessing.

Pretty sure that it won't exceed 20x the performance of upcoming R9 295X2 for a single GPU & even if it does it'd be only marginally higher than 20x, if at all, unless of course something like HSA takes off :awe:

 

[Stuka87]

Precisely between 0 and 5000% better. I'm sure by now you see there are no accurate ways of doing anything more than guessing.

So you are saying there is no possible way they will be slower? Hmm...

 

[Kenmitch]

So you are saying there is no possible way they will be slower? Hmm...

Maybe Haswel effect....Compacting transisters might not be good temp wise.

Guess it's possible to maybe spread the heat evenly in the design.

 

[Phynaz]

They will be better by 42.

 

[bunnyfubbles]

GPUs are redonkulously parallel in nature. With CPUs we've hit a wall because we're kind of stuck at a certain clock rate (ie ~5GHz) and its not easy to just increase IPC. The one thing we could be doing is cranking out a ton of more cores, but that's not exactly practical either as its difficult to leverage the extra cores in the many every day tasks that aren't inherently parallel like with what GPUs do.

So yeah, with GPUs we can expect performance scaling like we've been used to for a while longer now because we can just keep adding more "cores" as well as clock rate, as we've certainly not come close to any sort of clock rate wall for GPUs.

 

[Mondozei]

GPUs are redonkulously parallel in nature. With CPUs we've hit a wall because we're kind of stuck at a certain clock rate (ie ~5GHz) and its not easy to just increase IPC. The one thing we could be doing is cranking out a ton of more cores, but that's not exactly practical either as its difficult to leverage the extra cores in the many every day tasks that aren't inherently parallel like with what GPUs do.

So yeah, with GPUs we can expect performance scaling like we've been used to for a while longer now because we can just keep adding more "cores" as well as clock rate, as we've certainly not come close to any sort of clock rate wall for GPUs.

Thanks for a great comment, even if we don't know for certain we can make some educated guesses and set up a framework for thinking about the issue, like you have done, compared to some people in this thread who have little knowledge and post troll replies.

We are still to see 20nm GPUs. Then we get 20nm with finfets called 16nm later on. Then after that we may see a real 16nm.

Could you expand on that?

 

[dangerman1337]

I think Nvidia & AMD will kind of slow down with die processes, I won't be surprised if Nvidia's Volta & AMD's equivalent will be on 16nm FF but use efficiency improvements and stacked DRAM to increase performance per watt as the "next generation jump". We'll probably go from TSMC 16nm FF > TSMC 10nm for the next architectural leap after Nvidia's Volta & AMD's equivalent as it'll be more cost-efficient to do so by say 2018.

 

[ShintaiDK]

Could you expand on that?

Its marketing.

TSMC first releases 20nm, then 20nm with finfets. But to avoid looking so far behind as they are. They simply rename 20nm with finfets to 16nm. Global Foundries does the same.

 

[Railgun]

So you are saying there is no possible way they will be slower? Hmm...

Way to be ridiculously pedantic.

Ok...

The slowest new part will be slower than the fastest old part. That should satisfy your comment.

The speed will be the speed. The performance will be the performance. Its percentage improvement (or decline) will be whatever it ends up being, which is all speculative at best until it's not vaporware.