AMD 20nm GPU Discussion Thread

[TreVader]

Your analysis is entirely (and predictably) one sided. Everyone knows GK110 was designed for Nvidia's HPC ambitions first and foremost, hence the reason GK104, 106, and 107 have significantly higher perf/mm^2.

If you want to compare perf/mm^2 look at GCN 1.1 (either Bonaire or Hawaii) and compare it to Maxwell. It took AMD about the same amount of time to upgrade GCN 1.0 to 1.1 as it did for Nvidia to move from the first Kepler to Maxwell. GM107 has a smaller die and less transistors than Bonaire, but is faster and way more efficient. Nvidia is doing way more with less.

But if you want to stick with GK110 vs. Hawaii comparisons, consider power draw, heat, and all the other functions GK110 is capable of that Hawaii isn't. It is very obvious Hawaii is being pushed harder as the r290x than GK110 is as the gtx780ti. Also the exclusive compute specific features in GK110 don't come for free - die space and transistors are used. In fact, insofar as performance is concerned, stock gtx780ti vs. stockr290x is pretty similar to gtx580 vs. hd6970 (the previous node flagship cards). And when looking past that sole comparison, GK110 looks better in every regard vs. Hawaii than GF110 did vs. Cayman.

Bionare vs maxwell? So you pick the only two recent architectures that could possibly make nvidia look decent and want me to take that comparison as a prediction on 20nm? Ridiculous!


It's pretty simple… GK110 and hawaii are the respective high end 28nm parts, no? How is it even close to relevant to compare bionare to nvidias current newest chip? GK110 = 560mm. Hawaii = 435 mm. GK110 5.2 GFLOPS SP Hawaii = 5.7 GFLOPS SP

Hawaii = .0131 GFLOPs/mm^2
GK110 = .0093 GFLOPs/mm^2

If hawaii was 560mm^2 it would be 7.3 GFLOPs. That is how much better AMD is at compute, and I won't even mention, DP it gets ugly for nvidia there.


AMD purposely made hawaii a compute card, look at the W9100. If they want, they can grab the crown from nvidia any time they want.


As far as "bias", my only bias comes from owning both products for years. I think nvidia makes good products sometimes (6800,8800,285) but they always overcharge and they aren't as good with hardware as AMD. Do you own any AMD products right now?

 

[VulgarDisplay]

If Hawaii was scaled up to 560mm2 but kept the same transistor density it has at 435mm2 it would indeed be really fast. It would also probably be so hot that it would either require a triple slot cooler, or clocks would be reduced to keep heat in check.

 

[PPB]

You cant really compare their latest uarchs on the enthusiast/HPC segment because only 1 of both companies has a shipping product right now using that uarch at 28nm. Extrapolating 750ti and 7790 values wont do anything, as you lose perf/watt when you go for a bigger die.

If you want to do that, you need to do a rule of three involing 7790, 290x, and 750ti's performance per mm or per watt. I dont doubt that a temptative big maxwell die would come triumphant. Bolded for obvious reasons.

At the end, I think GCN 1.1 is NV's GK11x. We usually see 2 itinerations without big changes from NV, this being rather an exception because GK100 was canned and they really didnt show up signs of other GK11x products. Pirate Islands will probably more akin to Maxwell as both uarchs should bring more radical changes from their antecesors.

 

[VulgarDisplay]

More 4k benchmarks 290X vs 780ti with the fixed NV driver, the most recent WHQL release:

http://www.guru3d.com/articles_pages/asus_geforce_gtx_780_ti_matrix_review,16.html

and yet more:

http://www.guru3d.com/articles_pages/asus_geforce_gtx_780_ti_matrix_review,17.html

The 780ti has a commanding lead in all games tested @ 4k with the most recent WHQL. But the 290X is a better value since it is the cheaper card, and the mining inflation situation is starting to go away. And they're both fairly close in performance, and both very strong at 4k.

0-5 fps is now a commanding lead with undisclosed clocks. Is Hawaii in uber mode? The fact that the 290 is within 1-3 fps of the 290x sure makes me wonder.

 

[TreVader]

You cant really compare their latest uarchs on the enthusiast/HPC segment because only 1 of both companies has a shipping product right now using that uarch at 28nm. Extrapolating 750ti and 7790 values wont do anything, as you lose perf/watt when you go for a bigger die.

If you want to do that, you need to do a rule of three involing 7790, 290x, and 750ti's performance per mm or per watt. I dont doubt that a temptative big maxwell die would come triumphant. Bolded for obvious reasons.

At the end, I think GCN 1.1 is NV's GK11x. We usually see 2 itinerations without big changes from NV, this being rather an exception because GK100 was canned and they really didnt show up signs of other GK11x products. Pirate Islands will probably more akin to Maxwell as both uarchs should bring more radical changes from their antecesors.

I just did compare them, they are on the same node and represent each companies latest efforts.

You can't compare a 2-3yo design against a design that just came out today. NV and AMD don't synchronize their process changes, the only reason they're even on the same node is TSMC. GCN 1.1 is not nvidias anything, it's GCN 1.1. What you're doing is like comparing a Sandy bridge Pentium to an 8350 and then saying "obviously AMD has better performance, and that will continue to 20nm"

 

[TreVader]

0-5 fps is now a commanding lead with undisclosed clocks. Is Hawaii in uber mode? The fact that the 290 is within 1-3 fps of the 290x sure makes me wonder.

That's pretty bad. They don't even say if they used mantle! I think they just regurgitated their old hawaii 4k benchmarks and slapped on the new 4k for the 780ti.



Even on a rigged set, the 290X doesn't get trounced. And the 290X lead in 4k is likely to increase as more games come out with more than 3GB of VRAM use.

 

[Vesku]

The density and memory controller gains AMD made with Hawaii are more indicative of AMD next gen than specific GCN -> GCN 1.1 tweaks, imo. Taken out of the next gen toolbox and applied to 28nm.

 

[rtsurfer]

So when can we expect the AMD's 20nm performance equivalent of R9 290X.

My guess would be Q2 2015.

What do you guys think..??

 

[Bateluer]

If Hawaii was scaled up to 560mm2 but kept the same transistor density it has at 435mm2 it would indeed be really fast. It would also probably be so hot that it would either require a triple slot cooler, or clocks would be reduced to keep heat in check.

Which is why Hawaii is what it is on 28nm. The original plan was for 20nm, remember.

So when can we expect the AMD's 20nm performance equivalent of R9 290X.

My guess would be Q2 2015.

What do you guys think..??

Not unless they have serious problems with 20nm, bad tape outs and poor yields. They taped out in the last days of 2013, they'll be shipping at least some by the end of 2014 at the very latest. By Q2 2015, GloFlo's own 14/16nm process should be taping out its first silicon.

 

[rtsurfer]

Not unless they have serious problems with 20nm, bad tape outs and poor yields. They taped out in the last days of 2013, they'll be shipping at least some by the end of 2014 at the very latest. By Q2 2015, GloFlo's own 14/16nm process should be taping out its first silicon.

Usually with a new node poor yields do happen.

Plus if AMD is going to use stacked DRAM and all that fancy stuff, it might take some time.

Considering Hawaii came out end of October/ beginning if November 2013, Q2 2015 with a 1.5 year delay seems plausible to me.

 

[TreVader]

So when can we expect the AMD's 20nm performance equivalent of R9 290X.

My guess would be Q2 2015.

What do you guys think..??

Q4 2014 or Q1 2015. AMD usually puts out one big die to start, and activates all the features a la 290x/7970. Then later they produce higher clocked versions with extra vram.

NV usually puts out a smaller die, then re-tools for big die later. It's very possible GM104 will be out, then hawaii replacement, then GM110. It may or may not go in that order as far as performance too.

 

[Enigmoid]

Bionare vs maxwell? So you pick the only two recent architectures that could possibly make nvidia look decent and want me to take that comparison as a prediction on 20nm? Ridiculous!


It's pretty simple… GK110 and hawaii are the respective high end 28nm parts, no? How is it even close to relevant to compare bionare to nvidias current newest chip? GK110 = 560mm. Hawaii = 435 mm. GK110 5.2 GFLOPS SP Hawaii = 5.7 GFLOPS SP

Hawaii = .0131 GFLOPs/mm^2
GK110 = .0093 GFLOPs/mm^2

If hawaii was 560mm^2 it would be 7.3 GFLOPs. That is how much better AMD is at compute, and I won't even mention, DP it gets ugly for nvidia there.


AMD purposely made hawaii a compute card, look at the W9100. If they want, they can grab the crown from nvidia any time they want.


As far as "bias", my only bias comes from owning both products for years. I think nvidia makes good products sometimes (6800,8800,285) but they always overcharge and they aren't as good with hardware as AMD. Do you own any AMD products right now?

Well yeah, as a prediction for 20 nm the most recent architecture certainty makes sense for comparison.

The GK110 is likely optimized for lower power at the expense of a larger die. This is mainly for servers and professional markets where the cost of the die is made up by the larger margins.

GFLOPS is also a horrible way to measure performance, especially between architectures.

6970 = 2.7 TFLOPS
GTX 580 = 1.6 GFLOPS

Which was faster?

770 vs R280x is 3.2 TFLOPS vs 4TFLOPS.

Simply put it appears that nvidia is faster TFLOP for TFLOP than AMD.

 

[blackened23]

Which is why Hawaii is what it is on 28nm. The original plan was for 20nm, remember.

Just out of curiosity, do you have a good solid source on this? Because as far as I can tell, GPUs are in the planning phase for more than a year prior to launch. I don't think 20nm was in consideration given that timeframe, but I could be wrong. In any case, if you have a link to a good solid source to shed light on that.

I do know that cayman was affected by something similar, because 32nm was basically dropped by TSMC. But I haven't heard anything about Hawaii being planned for 20nm from the outset - as it was known for some time that 20nm would not be ready at any point in 2013. And that would definitely overlap with the planning and engineering phase for the GPU. My thoughts are that AMD knew from the get-go that 20nm would be an absolute no go. But if i'm wrong, a *good* source (ie, not semiaccurate or other national enquirer type tech news sites) would be great, i'm really curious about this.

In any case, it is unfortunate that GPU advancement is being held hostage somewhat by slowed advancement on the node process front. Really unfortunate. But on the other hand it does seem appreciable gains are still possible even at 28nm. Such as Apple's A7 which was a quantum leap over the prior A6. But surely chips of that calibur cost a ton of R+D and would be that much better on 20nm.

 

[sontin]

It's pretty simple… GK110 and hawaii are the respective high end 28nm parts, no? How is it even close to relevant to compare bionare to nvidias current newest chip? GK110 = 560mm. Hawaii = 435 mm. GK110 5.2 GFLOPS SP Hawaii = 5.7 GFLOPS SP

Hawaii = .0131 GFLOPs/mm^2
GK110 = .0093 GFLOPs/mm^2

If hawaii was 560mm^2 it would be 7.3 GFLOPs. That is how much better AMD is at compute, and I won't even mention, DP it gets ugly for nvidia there.

So, we ignore the simple fact that AMD can't hold their 1000MHz and need to throttle it down to <850MHz? :hmm:
GTX780TI uses 230W while running at 928MHz, the 290X only archives <850MHz with the same power.

 

[dangerman1337]

I think for AMD and Nvidia we'll see Performance Range (GM204 & R8 380(X)) and Mid range (GM206 & R7 370(X)) this year late Q3/Q4 and maybe the top range (GM200 & R9 390(X)) will be released early 2015 as games such as the Witcher 3, AC Unity will definitely and later on for the next two years will not have single 780tis running 2560x1440 at max/very high settings at a good frame-rate (if you have a 780ti at 1080p you will be likely fine).

 

[Cloudfire777]

Bionare vs maxwell? So you pick the only two recent architectures that could possibly make nvidia look decent and want me to take that comparison as a prediction on 20nm? Ridiculous!


It's pretty simple&#8230; GK110 and hawaii are the respective high end 28nm parts, no? How is it even close to relevant to compare bionare to nvidias current newest chip? GK110 = 560mm. Hawaii = 435 mm. GK110 5.2 GFLOPS SP Hawaii = 5.7 GFLOPS SP

Hawaii = .0131 GFLOPs/mm^2
GK110 = .0093 GFLOPs/mm^2

If hawaii was 560mm^2 it would be 7.3 GFLOPs. That is how much better AMD is at compute, and I won't even mention, DP it gets ugly for nvidia there.


AMD purposely made hawaii a compute card, look at the W9100. If they want, they can grab the crown from nvidia any time they want.


As far as "bias", my only bias comes from owning both products for years. I think nvidia makes good products sometimes (6800,8800,285) but they always overcharge and they aren't as good with hardware as AMD. Do you own any AMD products right now?

You do understand that AMD and Nvidia went to completely different routes here?

AMD concentrated on building a dense and hot chip with many transistors that give great GPGPU performance.
Nvidia concentrated on building a balanced chip that is good with gaming and GPGPU but also run very cool.

Clients are not just looking for the fastest workstation card. Many other factors comes to play like:
Is CUDA optimized for the programs and algorithms we use?
Do we have unlimited power and cooling supply for our workstations or do we need to get the best W/Performance and what about heat output?
How is the library and support for AMD cards and CUDA cards?

 

[tviceman]

In any case, it is unfortunate that GPU advancement is being held hostage somewhat by slowed advancement on the node process front. Really unfortunate. But on the other hand it does seem appreciable gains are still possible even at 28nm. Such as Apple's A7 which was a quantum leap over the prior A6. But surely chips of that calibur cost a ton of R+D and would be that much better on 20nm.

It is very discouraging, but at the same time it will force both Nvidia and AMD innovate.

 

[blackened23]

0-5 fps is now a commanding lead with undisclosed clocks. Is Hawaii in uber mode? The fact that the 290 is within 1-3 fps of the 290x sure makes me wonder.

You mean up to 10 fps? First of all, as you increase the resolution the performance graphs "compress" along with that increase in resolution. So in fact, yes, it is a commanding lead; many titles are 5-10 fps lead in favor of the 780ti compared to an aftermarket 290X (see the legitreviews link I posted earlier. They tested the Powercolor PCS triple slot card, which so far seems to be one of the fastest/best custom 290X cards).

So performance graphs get "compressed" as the resolution goes up. As an example, let's say you do benchmarks at 1080p. On average, maybe the 780ti has a 20-25 fps lead over the competition at 1080p. When you increase the resolution, the performance graphs compress a tad. At 2560x1600, the 20-25 1080p performance lead becomes 15 fps or so. At 4k? The lead becomes 5-10 fps. Yes, at 4k that is a commanding performance lead. And yes, while I don't think there's quiet mode on the Powercolor PCS custom 290X, I know it did not throttle or anything like that before you want to theorize such. It's an aftermarket custom card. View the benchmarks at 4k for yourself at legitreviews. Their powercolor PCS review is on the front page. No, it didn't throttle. No, it didn't use a clockspeed gimped quiet mode - to answer all of the questions brought up earlier.

That's not to say that the 290X is a slouch, it's clearly a very fast GPU and is cheaper. The 290X is a fast GPU in its own right and has more of the value and bang for the buck angle which AMD is generally known for. You likely won't be disappointed with either GPU at 4k resolution. My main point was the MST driver bug with NV has been resolved and has given Nv a significant 4k resolution performance bump. Obviously, a GPU being gimped at 70% GPU utilization will cause it to lose benchmarks - that was the old NV driver bug with MST, which has been fixed.

In any case, this is somewhat of a tangent meant to answer some of the questions in this thread. So with those questions answered, i'll leave that off topic alone, done talking about it. Anyway, about 20nm. If I may, the state of 4k resolution performance clearly indicates that 20nm may be the answer for that. So i'm pretty excited and optimistic for 20nm GPUs and Maxwell - they should be that much better for 4k resolution performance, perhaps at 20nm we can use a single GPU without having to compromise settings to get 60 fps in most titles. Of course you can do that now if you fiddle with settings a bit, but there are some out there who like doing ultra settings across the board. Maybe 20nm can bring us that. At 4k resolution. Without crossfire or SLI. One can hope. In any case, it will be interesting to see what happens with 20nm and next gen GPUs later this year.

 

[Keysplayr]

It is very discouraging, but at the same time it will force both Nvidia and AMD innovate.

Actually, I really like seeing what, to me, is a massive improvement architecturally on the same node. E.G. Maxwell GTX 750Ti.

Rather than to always MOSTLY rely on die shrinks to double the transistor count to "theoretically" double the performance of last gen, I like the fact that so much more can be done with what we have. Refinement is beautiful and an art form if you asked me. I know you didn't, but if you did.

 

[parvadomus]

Well yeah, as a prediction for 20 nm the most recent architecture certainty makes sense for comparison.

The GK110 is likely optimized for lower power at the expense of a larger die. This is mainly for servers and professional markets where the cost of the die is made up by the larger margins.

GFLOPS is also a horrible way to measure performance, especially between architectures.

6970 = 2.7 TFLOPS
GTX 580 = 1.6 GFLOPS

Which was faster?

770 vs R280x is 3.2 TFLOPS vs 4TFLOPS.

Simply put it appears that nvidia is faster TFLOP for TFLOP than AMD.

Your TFLOP for TFLOP logic is a bit flawed. First the 6970 was a VLIW based uarch a lot less flexible than the SIMD based uarch found in 580. That means that for the majority of compute tasks the 580 was able to stay closer to its theorical TFLOP rate.
And for the 770 vs 280, should I remember you that the 280 destroys the 770 on most tasks?

 

[desprado]

Your TFLOP for TFLOP logic is a bit flawed. First the 6970 was a VLIW based uarch a lot less flexible than the SIMD based uarch found in 580. That means that for the majority of compute tasks the 580 was able to stay closer to its theorical TFLOP rate.
And for the 770 vs 280, should I remember you that the 280 destroys the 770 on most tasks?

GTX 770 pawned R9 280x in gaming and benchmarking but not in mining.

 

[NTMBK]

GTX 770 pawned R9 280x in gaming and benchmarking but not in mining.

lol no. There are plenty of games where the R9 280X was much faster than the 770, and vice versa. For example, here's the 280X clearly winning:

But then switch games and what do you know:

It all depends on what games you play. It certainly didn't get "pwned", and at launch was priced $100 below the 770. (But then of course the damn mining craze happened.)

 

[parvadomus]

GTX 770 pawned R9 280x in gaming and benchmarking but not in mining.

I was talking about compute tasks, not gaming. In gaming that gpus are almost tied, and the specific game makes the difference.

 

[Fastx]

@NTMBK

Your comparing a aftermarket ASUS DCUII OC 280x to a stock clock reference GTX 770 in those BM charts, not a good/fair comparison imo.

 

[Enigmoid]

Your TFLOP for TFLOP logic is a bit flawed. First the 6970 was a VLIW based uarch a lot less flexible than the SIMD based uarch found in 580. That means that for the majority of compute tasks the 580 was able to stay closer to its theorical TFLOP rate.
And for the 770 vs 280, should I remember you that the 280 destroys the 770 on most tasks?

I was comparing SP not DP.

As for DP you care comparing theoreticals of 134 GFLOPS for the 770 and 1076 GFLOPS for the 7970 Ghz.

Anyway, needless to say different architectures can no be compared GFLOP for GFLOP.

That said Kepler (even GK 110) is weak DP.