New Zen microarchitecture details

[leoneazzurro]

By "at least" I meant that the total improvement should be at least that, since they claim it for the process alone. The improvements from the architecture leave at least something upon interpretation (GCN1/GCN2/GCN3), however the process statement does not.

What part of "up to" that is clearly stated on the slide also for the process is not clear to you?
Edit: I already stated there are reasons why the process of the GPU cannot give in every scenario or application the maximum improvement, especially if there are other bottlenecks in the CARD. in my previous post. Also not everything scales at the same level.Try to understand it.

 

[The Stilt]

My point is, as we have gone horribly off topic is that the 40% is likely a very best scenario case and we are highly unlikely to actually get that.

Based on AMD's previously quoted IPC improvements for a new µarch that is most likely the average improvement. Their quoted figures for Steamroller and Excavator were spot on with the average improvement. I would be equally shocked and disgusted if the average would be lower than that.

 

[The Stilt]

What part of "up to" that is clearly stated on the slide also for the process is not clear to you?
Edit: I already stated why the process cannot give always the maximum improvement in my previous post. Try to understand it.

Gimme a break. 2.35% <> 70%?

 

[leoneazzurro]

Gimme a break. 2.35% <> 70%?

2,35% in graphics workload is in the interval "up to" 70%, and in many tests (compute especiallY) it can be near the maximum.
Average is not peak.
Cards are not made only with a GPU.
Da you need some other evident statement or do you want to continue trolling?

 

[coercitiv]

Da you need some other evident information or do you want to continue trolling?

Welcome to the forums. Do you have any previous experience with The Stilt being a troll, or is this just a way to introduce yourself to the community?

My point is, as we have gone horribly off topic is that the 40% is likely a very best scenario case and we are highly unlikely to actually get that.

You know what, maybe this is indeed the healthy way to look at their claims. Zen is to have up to 40% increase in IPC, with a considerable chance of the average being way bellow this 40%.

Looking forward to Zen live demos of increased IPC if AMD wants to make the claim more credible.

PS: we're not that far OT considering the diminished perf/watt ratio is likely skewed by lower than expected performance of the 14nm process. This info, corroborated with the fact that AMD missed their power consumption target for the card (<150W) and the fact that their performance target was set by old gen cards (970, 390 etc - hence did not need adjusting on the fly) can only lead us to a single conclusion: AMD expected better results from their 14nm process. This affects Zen if true.

 

[leoneazzurro]

Welcome to the forums. Do you have any previous experience with The Stilt being a troll, or is this just a way to introduce yourself to the community?

Ty for teh welcome. Then explain how he is not trolling. We have cleared that "up to" is not "peak is always reached" especially when comparing systems with vastly different memory subsystem and especially not when comparing different workloads, continuing to repeat the same nonsense again and again how can it be classified?

 

[Abwx]

You know what, maybe this is indeed the healthy way to look at their claims. Zen is to have up to 40% increase in IPC, with a considerable chance of the average being way bellow this 40%.

Chances are that it will be higher than 40% on average, Zen is twice an EXV core for both management and exe ressources.

Or are you saying that with 100% more ALUs and 100%+ more FP ressources/core they will hardly get 40% more instructions/clock, in that case a 3 ALUs design, like SBridge and IB isnt it, along with a bigger cache would have been enough.

 

[leoneazzurro]

I found also the conditions specified in the slide. It was referring to a 470 vs the 270x

Compare the slides here https://www.computerbase.de/2016-06/radeon-rx-480-test/3/#bilder with the footnote here http://radeon.com/polaris-architecture/
(sorry but I did not find anything better in a short time)

It is also stated that "results are not average and may vary".

 

[coercitiv]

Or are you saying that with 100% more ALUs and 100% more FP ressources/core there will hardly get 40% more instructions/clock, in that case a 3 ALUs design along with a bigger cache would have been enough.

I'm saying for me personally the time to take AMD performance claims for granted is over. I'll believe it when I see a live demo.

 

[Abwx]

I'm saying for me personally the time to take AMD performance claims for granted is over. I'll believe it when I see a live demo.

In the meantime estimations should at least take account of what is known of the design, FTR, and as edited in my previous post, Zen has 33% more ALUs than SB and IB, yet we have people here doing speculation on IPC while blatantly, and likely willfully for some, ignoring this fact.

 

[The Stilt]

I found also the conditions specified in the slide. It was referring to a 470 vs the 270x

Compare the slides here https://www.computerbase.de/2016-06/radeon-rx-480-test/3/#bilder with the footnote here http://radeon.com/polaris-architecture/
(sorry but I did not find anything better in a short time)

It is also stated that "results are not average and may vary".

I'm fully aware that it is "up to", but still it is horrible marketing IMO. Would you personally be satisfied if your ordered an imaginary, expensive steak with beef content advertized as "up to" 100%, while in reality the "steak" was actually made of 99% of saw dust and 1% of actual beef?

One thing I know for certain, if I worked at AMD marketing dept. I wouldn't sleep too well.

Also, 37.7% (270X vs. RX 480, 1080-2160P) is still pretty far away of the advertized "up to" 180% (2.8x) improvement. I wonder in what kind of workload is the 180% improvement achievable, since obviously that figure has to be based on something :whiste:

 

[Abwx]

I wonder in what kind of workload is the 180% improvement achievable, since obviously that figure has to be based on something :whiste:

Count the TFlops.

That being said it s a Zen dedicated thread, dont see what GPUs have to do here.

 

[Doom2pro]

Count the TFlops.

That being said it s a Zen dedicated thread, dont see what GPUs have to do here.

I know right? Yesterday I had to double check to make sure I was about to post in the correct Forum...

I know they share a common Node but really? Every day?

 

[leoneazzurro]

I know it is marketing, but it is true for all producers to have their product look better than it is in reality (see also the nvidia slides posted before), and also not only in the semiconductor sector. If you buy an hybrid car, advertised as having low fuel consumption, and drive it mostly in highway for long journeys, you'll likely to be disappointed. But this is why all marketing must be taken with a grain of salt, and we need to be informed.

BTW, the footnotes, truth is always in the footnotes :-D

 

[The Stilt]

Count the TFlops.

Yeah, so it is close (2.73x) to the theoretical difference in GFLOP per watt figure. When the actual figure (power draw, 180W vs. ~120W actual) on 270X is used, the difference is ~1.85x.

Fair enough, I guess.

/Derail

 

[Abwx]

Yeah, so it is close (2.73x) to the theoretical difference in GFLOP per watt figure. When the actual figure (power draw, 180W vs. ~120W actual) on 270X is used, the difference is ~1.85x.

Fair enough, I guess.

/Derail

I wont pursue this useless debate but for your insight the 270X use 154W in Anno 2070 while the 480 is at about the same power but with more RAM at the rate of 20W/4GB, so you started with wrong numbers just to try demonstrating your indeed flawed conclusion.

http://www.hardware.fr/articles/908-5/bruit-temperatures-consommation.html

http://www.hardware.fr/articles/951-9/consommation-efficacite-energetique.html

 

[looncraz]

The consoles may not have brute force power, but the games look absolutely fantastic. That's a natural sacrifice between flexibility you get with PC, and efficiency with dedicated gaming devices like consoles, even if the hardware is identical.

Well, and lower resolutions (as you later mention) and only a 30FPS target... so, basically, the games are set to medium settings at 720p and only run at 30FPS... which you can configure a PC game to do without issues :biggrin:

All of that also reduces the bandwidth requirements compared to a 1080p card... 1080p requires 2.25x as much bandwidth for the frame buffer data.

 

[ElFenix]

I'm saying for me personally the time to take AMD performance claims for granted is over. I'll believe it when I see a live demo.

Has anyone ever? Not a one of us is actually committing to purchase (as in, nonrefundable deposit) one of these prior to actual benches.

 

[AtenRa]

Also, 37.7% (270X vs. RX 480, 1080-2160P) is still pretty far away of the advertized "up to" 180% (2.8x) improvement. I wonder in what kind of workload is the 180% improvement achievable, since obviously that figure has to be based on something :whiste:

According to AMD, 2.8X is between RX470 110W TDP vs R9 270X 180W TDP in 3D Mark Fire Stike.

Do the same for RX 480 (150W TDP) vs R9 290X (290W TDP) and I believe you will get to ~2.8x.

Edit, here so we can end this,

 

[looncraz]

Just so it's clear, the claim was made by Raja Koduri during the Computex presentation, we're not talking about some partial leak or out of context material. Those of you who haven't watched it but contribute to the conversation, please do take a look, it only takes 1 minute.

HOW have I not seen that!?! :'(

This makes perfect sense, then. 1.7x P/W improvement is easily being seen due to the process. We see closer to 2x on average because there is a 15% SP improvement on top of that. And the "With AMD technologies" part is talking about specific improvements in VR, tessellation, decode, compute, and everything else the GPU can do that doesn't always directly relate to FPS. Still, likely using a TDP vs TDP measurement.

 

[coercitiv]

Do the same for RX 480 (150W TDP) vs R9 290X (290W TDP) and I believe you will get to ~2.8x.

Using 3DMark scores only gets RX 480 to 2x perf/w over 290X.

It's also quite sad to come up with the idea we should compare with 290X when the prior gen card was 390X. (faster than 290X, 275W TDP)

 

[The Stilt]

I wont pursue this useless debate but for your insight the 270X use 154W in Anno 2070 while the 480 is at about the same power but with more RAM at the rate of 20W/4GB, so you started with wrong numbers just to try demonstrating your indeed flawed conclusion.

http://www.hardware.fr/articles/908-5/bruit-temperatures-consommation.html

http://www.hardware.fr/articles/951-9/consommation-efficacite-energetique.html

https://www.techpowerup.com/reviews/AMD/R9_270X/24.html

 

[looncraz]

Also AMD specifically claims that 1.7x of the total 2.8x comes from the 14nm LPP process transformation alone. So Polaris 10 should have AT LEAST 1.7x the performance per watt of ANY 28nm (even Fiji) GPU for the claim to be true. Since RX 480 has 2.35% higher performance per watt than Fiji (R9 Fury) according to TPU (1080 - 2160)...

https://www.techpowerup.com/reviews/AMD/RX_480/25.html

Nice job ignoring the difference in power usage of HBM vs GDDR5

We also have to ignore the super-low clocked Nano entirely... and the Fury X has the advantage of a liquid cooler keeping temps, and leakage, low.

In fact, a stock Fury X only pulls some 220W during gaming tests.

http://www.tomshardware.co.uk/amd-radeon-r9-fury-x,review-33235-7.html

Add 30W for RAM chips, another 15W for GDDR5 controllers, and another 15W for 80C temps, and we see 1.6969x reduced power... while clocking about 15~20% higher on average (albeit on a smaller GPU).

 

[Abwx]

https://www.techpowerup.com/reviews/AMD/R9_270X/24.html

You re insisting uselessly, Anno or the Witcher 3 loads GPUs way more than Crysis 2 (didnt TPU find something more ancient.??.), that s not by chance that they are used by Hardware.fr, besides they are closer to full throughput than Crysis 2 by the virtue of the higher power comsumption, i would have thought that comparisons should be done at rated throughput.

Edit : Last post from me about this OT debate...

 

[KTE]

You know what, maybe this is indeed the healthy way to look at their claims. Zen is to have up to 40% increase in IPC, with a considerable chance of the average being way bellow this 40%.

That's what I've been saying. IMHO, that's what it is.

We are going around in circles due to being starved of any proper info.

Every change mentioned to the uarch means nothing when you can't predict how much each will make a separate or combined difference to a full execution of different instructions.

That's why we are relegated to perusing in guesstimates ()