New Zen microarchitecture details

[The Stilt]

Actually I've seen quite a bit of variation in Hawaii ASIC quality as well, ranging from the high 60s to the low 80s. The Sapphire Vapor-X 290s I have are pretty awful.

I have a strong feeling that the leakage difference between SIDD 0 and SIDD 100 rating is much larger on 14nm LPP Polaris than it was on 28nm HPM Hawaii.

On Hawaii XT 0% SIDD = 16A, 100% = 26A. On Hawaii PRO (harvest SKU) the variation was much larger.

 

[itsmydamnation]

Eh?
I am talking about the 14nm process overall, I am not talking about power draw of the 480, that is an entirely different subject.
As I said, they can't talk about it right now, we have to wait for the SEC to release them, then, we can get a glimpse of what is going on by the documents they filed.

No you said based on 480 clocks, clocks for Zen doesn't look good.

Clocks and power consumption are parts of the same thing, they made architectural decisions based on hitting a perf/watt metric for mobile discrete for Polaris. Architectural decision will either enable or inhibit clock rate this will be just as much around target circuit timings as it is higher level architectural decisions.

The question that remains is how much different in perf/watt clock scaling could a AMD GPU be if targeted from group up to hit that ~150watt mark. We will have to wait until Vega to really see.

Why does everyone wanting to shit on Zen's clocks ( mostly because they can't find a uarch reason to do so) constantly ignore actual CPU's (mongoose and Kyro) clocks on LLP remember these cores are ~50% wider then CON/CAT / ~25% narrower then Zen.

Even A57 on LPE is 1.6watts a core @ 2.1ghz. A57 is significantly wider in execution then Zen because Zen has multiple RFP's and schedulers vs 57's shared scheduler/dispatch. While Zen is wider in Decode/retirement. A57's a ~15 stage pipeline for int, Zen will be longer then that ( i think around 20 stages) so even that will allow for higher clock ceiling vs a A57.

Serious, even stilt has said CON core says FMAX limited by the L2 ( its the horrible module cache design). Take a new cache system, take Excavators front end, take a new execution stage and take Excavators load/store stage. Its not magically going to become a clocking dog, people here are basically saying 14nm LPP is worse then 28nm bulk.............

 

[Doom2pro]

Where is everyone getting these Clock estimations for Zen? Do people seriously think GPU clocks are representative of CPU clocks?

Last I checked architecture is a pretty strong influence on clocks...

Seeing as how Zen is a higher IPC design than the speed demon Bulldozer based architectures, lower clocks (Relative to older architectures) are expected here.

If I had to totally guess here, I'd put low-high for Zen clocks at 2.4 to 3.2.

 

[Nothingness]

Even A57 on LPE is 1.6watts a core @ 2.1ghz. A57 is significantly wider in execution then Zen because Zen has multiple RFP's and schedulers vs 57's shared scheduler/dispatch. While Zen is wider in Decode/retirement. A57's a ~15 stage pipeline for int, Zen will be longer then that ( i think around 20 stages) so even that will allow for higher clock ceiling vs a A57.

A57 has separate issue queues and schedulers. The picture is misleading but the text in the box makes it clearer.

 

[CHADBOGA]

You can't expect people to click on a youtube link without some explanation of what they are going to see. I imagine most people just ignored your post. I did.

Same here.

 

[KTE]

No you said based on 480 clocks, clocks for Zen doesn't look good.

Clocks and power consumption are parts of the same thing, they made architectural decisions based on hitting a perf/watt metric for mobile discrete for Polaris. Architectural decision will either enable or inhibit clock rate this will be just as much around target circuit timings as it is higher level architectural decisions.

The question that remains is how much different in perf/watt clock scaling could a AMD GPU be if targeted from group up to hit that ~150watt mark. We will have to wait until Vega to really see.

Why does everyone wanting to shit on Zen's clocks ( mostly because they can't find a uarch reason to do so) constantly ignore actual CPU's (mongoose and Kyro) clocks on LLP remember these cores are ~50% wider then CON/CAT / ~25% narrower then Zen.

Even A57 on LPE is 1.6watts a core @ 2.1ghz. A57 is significantly wider in execution then Zen because Zen has multiple RFP's and schedulers vs 57's shared scheduler/dispatch. While Zen is wider in Decode/retirement. A57's a ~15 stage pipeline for int, Zen will be longer then that ( i think around 20 stages) so even that will allow for higher clock ceiling vs a A57.

Serious, even stilt has said CON core says FMAX limited by the L2 ( its the horrible module cache design). Take a new cache system, take Excavators front end, take a new execution stage and take Excavators load/store stage. Its not magically going to become a clocking dog, people here are basically saying 14nm LPP is worse then 28nm bulk.............

First, you talk about how clocks/power/architecture are all design set and linked.

Then you go on to contradict yourself boasting about very low power/clocking mobile CPUs using LPE/LLP and how that is somehow positively correlated to a high power/performance desktop CPU (Zen).

Capisce?

The two is like talking about how Atom is clocking@power using process A and positively deducing the same for Core.

Sent from HTC 10
(Opinions are own)

 

[DrMrLordX]

fwiw AMD has released a statement claiming that the 480's power problems are due to power regulation software. They claim that a driver fix is incoming. Interesting that they want to do it through the driver rather than card BIOS.

 

[Killrose]

Probably a bios for all new cards, software for people with cards to avoid a recall. Not everyone can bios flash.

 

[frozentundra123456]

Off topic in this thread, but AMD is shooting themselves in the foot with all these shoddy launches. Reminds me of the 290x debacle, only maybe worse. At least they had the driver right for that, they just released a card with a crappy cooler. This time they have both a crappy cooler and apparently a faulty bios. But lets hope for the sake of Zen it is only software, not the process itself.

 

[ZGR]

The thing that bothers me is the voltage being too high in nearly all of their products. Intel is guilty of the same thing, but nowhere near AMD's level. I understand the need for conservative voltages, but AMD is hurting their perf/watt in most of their products.

Will it be different for Zen?

 

[krumme]

The thing that bothers me is the voltage being too high in nearly all of their products. Intel is guilty of the same thing, but nowhere near AMD's level. I understand the need for conservative voltages, but AMD is hurting their perf/watt in most of their products.

Will it be different for Zen?

480 is mainstream product and yield must be high.
Going by SA numbers and random impression it seems lovering voltage and getting higher performance is possible for many. Now amd might have made a huge error here. Chance is they havnt

What is then left is imo more process variation than expected. To keep yields high on this mainstream card voltage is raised and vrm is programmed to fit reality. Efficiency and perf suffers. Remember it seems there is loads of cards going by eg Gibbos number.

It comes imo as no surprice the process is inmature and there is variation. Who would have predicted forhand gf could bring 480 to market at more or less same time as pascal on tsmc and in huge numbers?. Looking at history its incredible. Lol.

A metal respin extra for vega so to speak and better knowledge of process plus just a tad tad lower voltage for this highend card and it will fitt 250w tdp with perhaps 2.5 times perf.

Zen is targeted at servers and no way can they skimp on efficiency. 480 can take 20w extra no big deal but on zen the last thing we will see is imo high voltage and low efficiency. The entire business plan for the product will go down the drain.

When we had guessing contest a month ago i think i set the most conservative and low expectations for zen freq. But still i think it will be highly compettitive for its intended markets.

This situation changes nothing negatively. On the contrary. Gf seems to be ramping 14nm like crazy. We can just hope they have licenced 10nm and euv tech as well and is not trusting they can make it themselves with some ibm tech. It will end in disaster.

 

[Timmah!]

So what are the realistic performance expectations for Zen and when its supposed to be released? If i buy 6 core Broadwell CPU this summer, am i going to regret few months later, cause there will be 6/8 core Zen available with either better performance or price relative to performance or both? Or is the general expectation that AMD will at best match Intel at the same price-point?

 

[jpiniero]

480 is mainstream product and yield must be high.

GloFo doesn't really have any other customers using the SS14 node except for maybe Qualcomm. So I don't know if you can say that yield is high... if anything there's plenty of evidence that AMD loosened the bin quality of Polaris to get as many chips as they can even though there should be plenty of wafers available to them and they have to burn wafers anyway because of the WSA.

 

[krumme]

GloFo doesn't really have any other customers using the SS14 node except for maybe Qualcomm. So I don't know if you can say that yield is high... if anything there's plenty of evidence that AMD loosened the bin quality of Polaris to get as many chips as they can even though there should be plenty of wafers available to them and they have to burn wafers anyway because of the WSA.

Thats what i meant

 

[krumme]

So what are the realistic performance expectations for Zen and when its supposed to be released? If i buy 6 core Broadwell CPU this summer, am i going to regret few months later, cause there will be 6/8 core Zen available with either better performance or price relative to performance or both? Or is the general expectation that AMD will at best match Intel at the same price-point?

Q1 2017
95w tdp for 8 core
Cheaper than Intel for mb plus cpu because its from lesser brand and because of lower perf
Ib fpu ipc level
hsw integer ipc level
15% slower freq for 8 cores vs intel

Imo my guess. Nice lean and cost effective system if you want 8 cores and want it cheaper than intel with 80% the performance per core except for heavy fpu load. 50%- 60% cost. Vs intel cpu plus mb.

 

[scannall]

So what are the realistic performance expectations for Zen and when its supposed to be released? If i buy 6 core Broadwell CPU this summer, am i going to regret few months later, cause there will be 6/8 core Zen available with either better performance or price relative to performance or both? Or is the general expectation that AMD will at best match Intel at the same price-point?

Obviously nobody here knows the opening selling price or opening performance. While I expect it to be decent, with a lower price per core, we'll see.

I don't know about you, but when I build a system, I have a target budget in mind. If I can save 150-200 bucks on a MB and CPU for more than enough performance there, I can use that savings on a better GPU, SSD or whatever else I am targeting. It won't be long now, so patience. And when it's available run the numbers for your own build and see how it works out for you, and your purpose.

 

[swilli89]

Q1 2017
95w tdp for 8 core
Cheaper than Intel for mb plus cpu because its from lesser brand and because of lower perf
Ib fpu ipc level
hsw integer ipc level
15% slower freq for 8 cores vs intel

Imo my guess. Nice lean and cost effective system if you want 8 cores and want it cheaper than intel with 80% the performance per core except for heavy fpu load. 50%- 60% cost. Vs intel cpu plus mb.

Except for gamers I think this might actually end up being closer to 90-95% for highly threaded modern games.

 

[Rezist]

Except for gamers I think this might actually end up being closer to 90-95% for highly threaded modern games.

This what all gamers hope for to get system building costs down. If it doesn't happen that means stagnation for probably another 8 years.

 

[coffeemonster]

Q1 2017

last I heard they still had desktop ZEN planned for Q4 2016

 

[krumme]

last I heard they still had desktop ZEN planned for Q4 2016

Look at ref 480 fmax. 480 was q2.
If they force zen out the door even if process hindering efficiency and freq and dont take another metal spin if needed - after 6/8 years of being behind - they simply dont have the the robust topmanagement needed. Then they are weak.

 

[KTE]

"The obvious question to ask is this - what if AMD's strategy for the mainstream is a product of necessity based on the limitations it has encountered in dealing with the new, immature 14nm FinFET fabrication process? In short, is AMD's value proposition the best possible option it had available based on the quality of technology it could actually bring to market?"

http://www.eurogamer.net/articles/d...-graphics-strategy-lead-to-mainstream-success

My thoughts, pretty much. The process is not looking good... just what I feared. Hello Phenom. This is a Q1 '17 product, no question.

Sent from HTC 10
(Opinions are own)

 

[PliotronX]

My thoughts, pretty much. The process is not looking good... just what I feared. Hello Phenom. This is a Q1 '17 product, no question.

Sent from HTC 10
(Opinions are own)

They're obviously not going to beat Intel at their own game but if they end up in the same ballpark using SMT at a more sane cost than Intel equivalents, it has a place in the market. Phenom was more like Phe-bomb but Zen will be the first time we might see AMD parts with performance greater than Core 2 Quad. It'll be interesting either way. More interesting than Intels 5% every generation.

 

[KTE]

The thing is, basic physics dictates they use a lower operating voltage for the FETs if the process allows (leakage remains under control). One of the main aims of scaling infact was precisely this, allowing for a quadratic fall in power consumption with scaling allowed to keep power density constant (what Dennards Law also overlooked).

If you don't see this happening, there's already 'problems'.

Edit: Intel went from 12 to 16 GPU EUs, 250mill more transistors and still lost 20W in max power (2700K 32nm to 3700K 22nm), due to finFETs.

Sent from HTC 10
(Opinions are own)

 

[senseamp]

"The obvious question to ask is this - what if AMD's strategy for the mainstream is a product of necessity based on the limitations it has encountered in dealing with the new, immature 14nm FinFET fabrication process? In short, is AMD's value proposition the best possible option it had available based on the quality of technology it could actually bring to market?"

http://www.eurogamer.net/articles/d...-graphics-strategy-lead-to-mainstream-success

My thoughts, pretty much. The process is not looking good... just what I feared. Hello Phenom. This is a Q1 '17 product, no question.

Sent from HTC 10
(Opinions are own)

Why do people assume that if AMD is inferior in some aspect, it's the fab's fault, and AMD did the best it could with what it got?
How's 14ff an immature process if Samsung and Apple have been shipping chips with it in high volume for over a year now?
The only thing that seems immature is AMD's design methodology, which effectively make their 14nm chips look like competition's 28nm chips from perf/watt perspective. So they are basically two process nodes behind Intel, one actual node due to Intel's process superiority, and one effective node thanks to AMD's methodology inferiority. They need for their methodologies to compensate for their process inferiority, but instead they are compounding it.
Making efficient high performance chips is not just about the process scaling, it's also about design, and there is an increased competition for good chip designers. Even Google is designing their own chips now. AMD has a weak position in this competition for talent, so why do people expect that of all companies, it's actually AMD's designers who are doing the best that anyone can do, and the sub-par results are all the process' fault?
I would temper Zen expectations with this in mind. Maybe it's great, but I wouldn't bet on it. AMD has a track record of over-promising and under-delivering.

 

[The Stilt]

Aging (maturity) of the process can change things drastically. The 32nm SHP SOI is one of the best examples. It started as a complete lemon with some huge issues with the leakage and the general inconsistency. In it's current form the 32nm SHP SOI is a completely different animal. Most likely one of the best processes ever made. The silicon consistency is immaculate and the average static leakage on full Vishera dies is up to 45% lower than it was at the time of launch.

The only trouble is that it took around three years for it to reach it's peak D:

So the 14nm LPP will most likely still improve heavily as the machinery and the process itself grow older, however the only question is will the improvement be enough. If Zeppelin would ship on as immature process as Polaris 10 did, it would be a complete disaster. Zeppelin is obviously somewhat smaller than Polaris 10, but not by a huge margin. Personally I would rather see Zeppelin to be delayed slightly than to be released on the 14nm LPP in it's current state.