New Zen microarchitecture details

[krumme]

SKL at 4.8GHz is 24% faster than SNB at 4.8GHz in the test you linked.

818 vs 1007?

 

[Arachnotronic]

818 vs 1007?

(1007-818)/818 = 0.231. I meant to write 23%.

 

[DrMrLordX]

Consumers don't know that and don't care, including me. AMD has been saying, "We're back! Here comes Zen! Faster than Broadwell-E! Desktop enthusiast chip inbound!". If the thing gets stuck at 3ghz with Ivy IPC, its a fail, even with 8 cores. Someone who wants 8 slow Ivy cores could have already had 6 fast Ivy cores that perform the same or better for several years already. This would be too ridiculous to be true. It goes north of 4Ghz or bust.
Lets not pretend like AMD isn't positioning this as a high end, enthusiast desktop CPU, because they are. If its slower than an OC'd Sandy, they are finished in the eyes of enthusiasts and gamers and many more people as well. They can't blow it that big this time, not after Bulldozer. There's just no way. They won't recover from it ever. It will be the death blow and they know it. They'd be better off avoiding any sales on the desktop altogether.

Three things:

1). 40% ST IPC over XV isn't that great. All the CMT designs have poor ST IPC and make up for it with the second thread. Switching to an SMT design *similar to Intel's* should have changed the power balance so that more of the Zen core's resources could be committed to the "first thread" where necessary. If Keller went with something closer to POWER8's SMT design then maybe we're just seeing the next evolution of Con cores.

2). At least according to one pre-release benchmark, it looks like Zen's throughput is slightly higher than Broadwell-E in an apparently-non AVX benchmark. Which is a step in the right direction, since even 8m/16t XV would probably have struggled in that department. What that says about Summit Ridge's "IPC" is anyone's guess. Is it a throughput monster like a Con chip? I just don't know. Everything we've seen about the dynamic nature of the core tells me that it should be able to commit most of its resources to a single thread. And that should result in ST IPC much higher than 40% over XV in certain scenarios.

3). I am skeptical of the claims of low launch speeds, but we'll see. Bear in mind that we enthusiasts may be able to fix the problem with overclocking. Lots of people are running out to buy Broadwell-E for their elite desktops right now and are struggling to OC past 4.4 GHz. It will be interesting to see how far we can push Summit Ridge. Maybe not that far, though . . .

 

[krumme]

(1007-818)/818 = 0.231. I meant to write 23%.

Yeaa. 23% for cb
What is a good bm of Integer perf improvement? And what is the difference here?

 

[raghu78]

Consumers don't know that and don't care, including me. AMD has been saying, "We're back! Here comes Zen! Faster than Broadwell-E! Desktop enthusiast chip inbound!". If the thing gets stuck at 3ghz with Ivy IPC, its a fail, even with 8 cores. Someone who wants 8 slow Ivy cores could have already had 6 fast Ivy cores that perform the same or better for several years already. This would be too ridiculous to be true. It goes north of 4Ghz or bust.
Lets not pretend like AMD isn't positioning this as a high end, enthusiast desktop CPU, because they are. If its slower than an OC'd Sandy, they are finished in the eyes of enthusiasts and gamers and many more people as well. They can't blow it that big this time, not after Bulldozer. There's just no way. They won't recover from it ever. It will be the death blow and they know it. They'd be better off avoiding any sales on the desktop altogether.

Do not get carried away by AMD marketing & PR. We saw how Polaris was hyped in terms of perf/watt by AMD and where it landed. AMD just flat out lied about their product. Polaris struggles to even match Maxwell in perf/watt. Given the fact that Zen is built on a high density 14nm FINFET process I would say a max clock even after OC would be 3.5-3.6 Ghz. Stock clocks will also be low. I would guess we might see 3 Ghz base / 3.4 Ghz in a best case scenario. Carrizo which was built on 28nm high density process maxed out at around 3.6 Ghz - 3.8 Ghz. I think Zen will also do more or less the same. Maybe in early 2018 with Zen+ we might see 4+ Ghz clocks as we are seeing with Bristol Ridge. I also do not expect AMD to match Intel in perf/watt. The Polaris marketing & PR hype has left me doubting anything AMD says. In fact I think the opposite of their claims is true.

 

[leoneazzurro]

Do not get carried away by AMD marketing & PR. We saw how Polaris was hyped in terms of perf/watt by AMD and where it landed. AMD just flat out lied about their product. Polaris struggles to even match Maxwell in perf/watt. Given the fact that Zen is built on a high density 14nm FINFET process I would say a max clock even after OC would be 3.5-3.6 Ghz. Stock clocks will also be low. I would guess we might see 3 Ghz base / 3.4 Ghz in a best case scenario. Carrizo which was built on 28nm high density process maxed out at around 3.6 Ghz - 3.8 Ghz. I think Zen will also do more or less the same. Maybe in early 2018 with Zen+ we might see 4+ Ghz clocks as we are seeing with Bristol Ridge. I also do not expect AMD to match Intel in perf/watt. The Polaris marketing & PR hype has left me doubting anything AMD says. In fact I think the opposite of their claims is true.

While I don't like marketing in general, I have to say that in this case you are a little mislead. As told several times, they did not present a comparison to Nvidia consumption, they compared Polaris with the Hawaii/Tonga generation, and it's quite clear that in this comparison the figures could be very near to what AMD declared. Also, according to AT very tests:

http://www.anandtech.com/show/10446/the-amd-radeon-rx-480-preview/6

RX480 has a quite better perf/W ratio than the i.e. the GTX970 (Maxwell based), not even speaking about the 470 and also the 460 fares quite well comparing the 960. Of course, Pascal based cards are still miles better in that regard, but your statement is incorrect. What AMD marketing said can be perceptively misleading if you don't take the words for what literally they say.
What AMD says for now, is that in that particular test, Zen is working better than an equally clocked 8-core Broadwell. Which can be. Now, taking it with a grain of salt, thay did not declare that with any workload Zen will perform better than Broadwell of even SB/IB, what they declared is that Zen will have an IPC improved at least 40% from XV and that in that very multithreading test it performed better.
All the rest is speculation/illusion/delusion/prejudice/expectation until actual third party tests come out, in the good and in the bad.
My personal expectation is that they will have a better gaming behavior compared to XV but still not on par with Intel offerings, while presenting their eight core product in direct competition to the Quad core parts from Intel, thus performing better in heavily threaded/FP intensive applications, but of course I acknowledge I can be totally wrong, in the good or in the bad side of the things.

 

[KTE]

IMO sub-3.4GHz base for the top model is a product failure. It will be too incompetent in today's landscape versus KBL and CNL. That's a product for 2014.

Sent from HTC 10
(Opinions are own)

 

[Lepton87]

is that Zen will have an IPC improved at least 40% from XV and that in that very multithreading test it performed better.
All the rest is speculation/illusion/delusion/prejudice/expectation until actual third party tests come out, in the good and in the bad.
My personal expectation is that they will have a better gaming behavior compared to XV but still not on par with Intel offerings, while presenting their eight core product in direct competition to the Quad core parts from Intel, thus performing better in heavily threaded/FP intensive applications, but of course I acknowledge I can be totally wrong, in the good or in the bad side of the things.

I wouldn't presume at least 40% I'd say they meant 40% on average.

IMO sub-3.4GHz base for the top model is a product failure. It will be too incompetent in today's landscape versus KBL and CNL. That's a product for 2014.

It seems to me that 8-core ZEN will be competing with Intel HEDT offerings not in price but for the same usage. Everything suggests that heavy-threaded workloads are the target for ZEN. Zen designs goals are similar to Intel's HEDT CPUs where throughput is just as important as single threaded performance. Of course it will be a disappointment for a market like gaming where Intel targets their high-clocking quad core chips on a newest architecture. Even intel HEDT offerings are worse for this usage model.

 

[raghu78]

While I don't like marketing in general, I have to say that in this case you are a little mislead. As told several times, they did not present a comparison to Nvidia consumption, they compared Polaris with the Hawaii/Tonga generation, and it's quite clear that in this comparison the figures could be very near to what AMD declared. Also, according to AT very tests:

http://www.anandtech.com/show/10446/the-amd-radeon-rx-480-preview/6

RX480 has a quite better perf/W ratio than the i.e. the GTX970 (Maxwell based), not even speaking about the 470 and also the 460 fares quite well comparing the 960. Of course, Pascal based cards are still miles better in that regard, but your statement is incorrect. What AMD marketing said can be perceptively misleading if you don't take the words for what literally they say.
What AMD says for now, is that in that particular test, Zen is working better than an equally clocked 8-core Broadwell. Which can be. Now, taking it with a grain of salt, thay did not declare that with any workload Zen will perform better than Broadwell of even SB/IB, what they declared is that Zen will have an IPC improved at least 40% from XV and that in that very multithreading test it performed better.
All the rest is speculation/illusion/delusion/prejudice/expectation until actual third party tests come out, in the good and in the bad.
My personal expectation is that they will have a better gaming behavior compared to XV but still not on par with Intel offerings, while presenting their eight core product in direct competition to the Quad core parts from Intel, thus performing better in heavily threaded/FP intensive applications, but of course I acknowledge I can be totally wrong, in the good or in the bad side of the things.

AMD started making claims of 2x perf/watt vs Maxwell when they demoed Polaris in early 2016.

http://www.anandtech.com/show/9886/amd-reveals-polaris-gpu-architecture

The reality was they struggled to even match the Maxwell parts in perf/watt.

https://www.techpowerup.com/reviews/ASUS/RX_460_STRIX_OC/25.html

AMD's marketing & PR flat out lied about Polaris efficiency. So I do not believe anything that AMD claims anymore. imo AMD Summit Ridge will struggle to match Intel Broadwell-E in absolute performance and perf/watt.

 

[Shamrock]

My expectations from post #1259 still stand:

I don't take this article as a gospel, but they claim, after the 3Ghz demo against Intel, that it will ship with higher frequency than the 3Ghz.

http://venturebeat.com/2016/08/18/amds-takes-biggest-jab-at-intel-in-years-with-zen-processor/

AMD said that power would be competitive, the frequency we saw would be even higher at production than what we saw, and production of what we saw could be produced at scale.

 

[LTC8K6]

Power would be "competitive"?

That actually doesn't sound great.

 

[moonbogg]

I think 8 core Zen will surprise people. I think it will be clocked decently and will OC to an acceptable degree, offering more value than expected.

 

[KTE]

I don't take this article as a gospel, but they claim, after the 3Ghz demo against Intel, that it will ship with higher frequency than the 3Ghz.

http://venturebeat.com/2016/08/18/amds-takes-biggest-jab-at-intel-in-years-with-zen-processor/

That can mean 3.2GHz...

Sent from HTC 10
(Opinions are own)

 

[AtenRa]

Just so I can understand, whats the problem with ZEN 8Core 16Threads 3.2GHz CPU at 95W TDP ??

Intel Core i7 9600K is 8Core 16Threads 3.2GHz base 3.7 Turbo (Single Core) at 140W TDP.

Are we 100% sure this 8Core 16Thread ZEN will only be 95W TDP ???

 

[moonbogg]

We won't see it for another 6 months anyway. We got plenty of time.

 

[leoneazzurro]

AMD started making claims of 2x perf/watt vs Maxwell when they demoed Polaris in early 2016.

http://www.anandtech.com/show/9886/amd-reveals-polaris-gpu-architecture

The reality was they struggled to even match the Maxwell parts in perf/watt.

https://www.techpowerup.com/reviews/ASUS/RX_460_STRIX_OC/25.html

AMD's marketing & PR flat out lied about Polaris efficiency. So I do not believe anything that AMD claims anymore. imo AMD Summit Ridge will struggle to match Intel Broadwell-E in absolute performance and perf/watt.

As I showed you before, it is not true that Polaris struggles, you took the worst review in the new regarding the power consumption, because TPU suite has a lot of very old games (I wonder if there is still someone playing some of them today) in the suite which just ran better on NV's architecture. Most of other reviews in the net suggest otherwise, as in the very own AT review, especially in comparison with the 970/980.
Also, the claim was that, in that very game, at the same frame rate, compared to a GTX950, the rx460 ran at 60 fps with 2x lower consumption. They never stated that every Polaris card would ran at half power consumption than every Maxwell card.
Which is quite feasible, especially if the chip (which is also geared for the mobile market) can support power adjustement according to the frame rate, which is also supported by NV chips, and they compared the 460 to the competing (in the price bracket, that is) card.
See, if you say they lie, it is not true. AFAIK they lied one time (or better, not marketing itself, but mostly an AMD employee that gave misleading "preview" information) and that was with Bulldozer.
The other times, they did not lie. They present their parts emphasizing the "pro" and hiding the "cons" but this is what every marketing does. If a customer has expactations based on what a marketing (whatever firm is not important) says then it's his problem. All marketing and declared data are lies on your term. Look at other markets, i.e. cars. Every manufacturer has to declare fuel consumption, but when do you in reality achieve what it is declared. Plainly said, never.
If you say "I will never believe to what AMD says", good. You should have never believed in first place. Like you should never straightly believe to Intel, or Nvidia, or every other manufacturer.

 

[ShintaiDK]

AMD started making claims of 2x perf/watt vs Maxwell when they demoed Polaris in early 2016.

http://www.anandtech.com/show/9886/amd-reveals-polaris-gpu-architecture

The reality was they struggled to even match the Maxwell parts in perf/watt.

https://www.techpowerup.com/reviews/ASUS/RX_460_STRIX_OC/25.html

AMD's marketing & PR flat out lied about Polaris efficiency. So I do not believe anything that AMD claims anymore. imo AMD Summit Ridge will struggle to match Intel Broadwell-E in absolute performance and perf/watt.

If AMD had something good, they show more of it rather than a canned benchmark in a secret setting. So you are spot on. This is the same as their Polaris setup in February. We all know how that turned out.

 

[Abwx]

If AMD had something good, they show more of it rather than a canned benchmark in a secret setting. So you are spot on. This is the same as their Polaris setup in February. We all know how that turned out.

You should read what AMD stated at their demo and at Hot Chips rather than making up theories..

With Blender Zen has better perf/watt at 3GHz than the 6900K at the same frequency, so it will be better at 3.2GHz as well, not counting that AMD surely ran the i7 at stock voltage and this latter is more a fraud than anything else since Intel delivered chips that are undervolted below qualification specs of consumer products...

 

[KTE]

If AMD had something good, they show more of it rather than a canned benchmark in a secret setting. So you are spot on. This is the same as their Polaris setup in February. We all know how that turned out.

I'm with you on this... If nothing more in detail is shown by Oct-Nov.

Sent from HTC 10
(Opinions are own)

 

[ShintaiDK]

You should read what AMD stated at their demo and at Hot Chips rather than making up theories..

With Blender Zen has better perf/watt at 3GHz than the 6900K at the same frequency, so it will be better at 3.2GHz as well, not counting that AMD surely ran the i7 at stock voltage and this latter is more a fraud than anything else since Intel delivered chips that are undervolted below qualification specs of consumer products...

And what are you doing now?.

How did your 14LPP claims work out for Polaris?

 

[Abwx]

And what are you doing now?.

How did your 14LPP claims work out for Polaris?

No theories from me since i use available datas, it s not like AMD didnt give a few hints, they not only stated that it consume lesss than the 6900K, they also gave a comparison with their current 28nm core..

https://www.computerbase.de/2016-08/amd-zen-architektur/

What does this mean according to your theories..?.

 

[leoneazzurro]

Polaris turned up in a competitive product. If someone expected a GP100 cruching chip at 249$ it's their issue. Also, the jump to 14nm had a very positive effect on power, if you compare their previous architecture at 28nm (and there, the process was the same as Nvidia's, so it's clear that the current GCN's chips are not easily power gated as the competitor's and it's not a process fault)

 

[ShintaiDK]

No theories from me since i use available datas, it s not like AMD didnt give a few hints, they not only stated that it consume lesss than the 6900K, they also gave a comparison with their current 28nm core..

https://www.computerbase.de/2016-08/amd-zen-architektur/

What does this mean according to your theories..?.

What data? Where does it show that Intel chips are undervolted?

AMD also claimed 2.8x perf/watt for Polaris, just to example the latest one.

40% wouldn't even put it close to Broadwell either if you compare with Excavator.

 

[leoneazzurro]

They claimed "up to 2,8x" compared to their previous 28nm chips, and AFAIK they delivered, roughly.
But I see that fanboyism is occurring on both sides (the claim of "undervolted Intel chips" makes me laugh, too), good debate about nothing.

 

[deasd]

Polaris turned up in a competitive product. If someone expected a GP100 cruching chip at 249$ it's their issue. Also, the jump to 14nm had a very positive effect on power, if you compare their previous architecture at 28nm (and there, the process was the same as Nvidia's, so it's clear that the current GCN's chips are not easily power gated as the competitor's and it's not a process fault)

Yes, 2.8Ghz based ES pretty fits my expectation for 3.0-3.2Ghz based for retail, and this is true benefit from process shrinking from 28nm to14nm. This is a huge leap from BD family, and it's interesting to see it against similar config from intel.