Zen 2 samples are with AMD, pumping away with 8C/16T at 4.5GHz .. TweakTown

[Thunder 57]

* More fully symmetrical ALU's to support more branch/SMO/CRC operations as Zen already having PLENTY of power for floating point calculations (except AVX niché)

* Enhanced Stack Engine to reduce the need of the quite already busy AGU's in Zen. That should knock of some bottlenecks. I find AMD going with 3 AGU's unlikely in Zen2.

I do wonder how much of a bottleneck the AGU's are. I doubt we see a third either, but I think it will be necessary. Also, can you really add any more ALU's while keeping just two AGU's?

When Ryzen was released I read that Zen 2 is a significant improvement beyond the die shrink of the CPU. It's supposed to be far more significant than Zen+ was. So an IPC improvement well beyond 15% is highly likely.

Doubt it. 15% is probably about the max we can expect.

 

[Glo.]

Zen 2 should be 2-5% faster, IPC wise, than Skylake/Kaby Lake/Coffee Lake.

 

[epsilon84]

Zen 2 should be 2-5% faster, IPC wise, than Skylake/Kaby Lake/Coffee Lake.

What do you base this on?

To achieve that would require a 10 - 15% increase in IPC, 15 - 20% if you go by Stilts testing which puts CFL approx 14% ahead of Zen 2 (12nm refresh)

Not saying its impossible, but I'll believe it when I see it. Would be good to finally have a uarch that beats 2015 era IPC...

 

[EXCellR8]

Is Zen 2 AM4+ or AM5... or ???

Sucks having to buy more new boards all the time but w/e.

 

[PotatoWithEarsOnSide]

AM4

Zen, Zen+, Zen 2, and Zen 2+ are all AM4, then likely a new socket for Zen 3.

 

[ao_ika_red]

AM4

Zen, Zen+, Zen 2, and Zen 2+ are all AM4, then likely a new socket for Zen 3.

Don't forget that dreaded Bristol Ridge is also on AM4 platform.

 

[Abwx]

What do you base this on?

To achieve that would require a 10 - 15% increase in IPC, 15 - 20% if you go by Stilts testing which puts CFL approx 14% ahead of Zen 2 (12nm refresh)

Not saying its impossible, but I'll believe it when I see it. Would be good to finally have a uarch that beats 2015 era IPC...

I suggest that you check the review below and compare the 2600X to the 8700K and see if there s as much as you state...

https://forums.anandtech.com/thread...0ti-computerbase.2554529/page-5#post-39596565

Other than this they announced 10-15% surely in respect of Summit Ridge, or 7-11% re Pinnacle Ridge, also Lisa Su stated that Zen 2 will be much better in games, albeit in this case there are other factors at work than brute IPC numbers.

 

[epsilon84]

I suggest that you check the review below and compare the 2600X to the 8700K and see if there s as much as you state...

https://forums.anandtech.com/thread...0ti-computerbase.2554529/page-5#post-39596565

Other than this they announced 10-15% surely in respect of Summit Ridge, or 7-11% re Pinnacle Ridge, also Lisa Su stated that Zen 2 will be much better in games, albeit in this case there are other factors at work than brute IPC numbers.

This 'IPC' debate has been done to death already, there is no 'universal' IPC measurement, the range I've seen is anywhere from 7% - 14% depending on the workload and how much AVX software is included in the testing suite - using more will obviously skew the results towards Intel.

To exceed Intel by 2 - 5% would require at least a 10 - 15% increase on average and I stand by that, perhaps more in gaming code, and I'm hoping you (and Lisa) are right in that Zen 2 will be much better in games, because that is the main reason I upgrade my CPU these days, would be great to have something faster than my 8700K (or possibly 9900K) next year as I plan to upgrade my GPU then as well.

 

[Paul98]

I am looking forward to this, it will be a good drop in upgrade for my 1700. Hopefully prices for a good GPU upgrade will be reasonable

 

[DrMrLordX]

Zen2's final performance may be heavily influenced by supported DDR4 speeds. Needs moar IF speed pl0x.

Don't forget that dreaded Bristol Ridge is also on AM4 platform.

The only people dreading Bristol Ridge are the ones using it.

 

[Abwx]

This 'IPC' debate has been done to death already, there is no 'universal' IPC measurement, the range I've seen is anywhere from 7% - 14% depending on the workload and how much AVX software is included in the testing suite - using more will obviously skew the results towards Intel.



To exceed Intel by 2 - 5% would require at least a 10 - 15% increase on average and I stand by that, perhaps more in gaming code, and I'm hoping you (and Lisa) are right in that Zen 2 will be much better in games, because that is the main reason I upgrade my CPU these days, would be great to have something faster than my 8700K (or possibly 9900K) next year as I plan to upgrade my GPU then as well.

Whenever i look at about any review the difference between a 2600X and a 8700K is about 11% despite an advantage of roughly 350Mhz all core turbo, so i have no doubt that Zen 2 could be well ahead of CFL.

https://www.hardware.fr/articles/975-17/indices-performance.html

Also, if out of 20 benchmark 19 are at even scores and there s a single one where the difference is 100%, for whatever reason, then the average will be 5% better IPC, you could as well do so with 18 even scores and 2 that show a 50% advantage, that was basicaly Computerbase.de methodology (with 3DPM V1.0 and Euler 3D) as well as The Stilt s.

 

[slashy16]

4.5ghz? so AMD is about to match frequencies Intel shipped 5 years ago. Intel is shipping 5ghz parts and it appears they could easily release 5.2-5.3ghz parts on 14nm. Intel's 10nm delays are horrible but, don't seem to be as bad now. It looks like they will remain competitive for the remainder of 2018/2019

 

[scannall]

4.5ghz? so AMD is about to match frequencies Intel shipped 5 years ago. Intel is shipping 5ghz parts and it appears they could easily release 5.2-5.3ghz parts on 14nm. Intel's 10nm delays are horrible but, don't seem to be as bad now. It looks like they will remain competitive for the remainder of 2018/2019

An early engineering sample on an engineering board doesn't really say much one way or the other about the speeds the parts will ship with. ;-) What speeds will it actually ship with? I have no idea. And any AMD employees who read this board aren't going to say anything. So we'll have to see when they are released and tested.

 

[french toast]

4.5ghz? so AMD is about to match frequencies Intel shipped 5 years ago. Intel is shipping 5ghz parts and it appears they could easily release 5.2-5.3ghz parts on 14nm. Intel's 10nm delays are horrible but, don't seem to be as bad now. It looks like they will remain competitive for the remainder of 2018/2019

Remember we heard of early zen engineering samples reaching 2.8ghz and 3.0ghz with sandy bridge PPC...one year later we got the 1800x!.

I think it is encouraging to see 4.5ghz this early on a buggy engineering sample.
If they could squeeze out 4.7-4.8ghz with ~15% PPC then that would be awesome imo.

 

[Lovec1990]

but question is 4.5Ghz single core or all cores

 

[AtenRa]

I would be very pleased with 2700X performance at 45W TDP.

 

[Zucker2k]

Remember we heard of early zen engineering samples reaching 2.8ghz and 3.0ghz with sandy bridge PPC...one year later we got the 1800x!.

I think it is encouraging to see 4.5ghz this early on a buggy engineering sample.
If they could squeeze out 4.7-4.8ghz with ~15% PPC then that would be awesome imo.

On a TSMC process? Has that ever happened with any of their processes? Even Global Foundries, who I believe were more focused on high performance nodes (Bulldozer, for example) couldn't keep up with Intel in the GHz race. No doubt, these upcoming chips will improve upon zen+ but something about TSMC and high 4.8GHz makes me real skeptical.

 

[coercitiv]

On a TSMC process? Has that ever happened with any of their processes?

Invoking historical data isn't really a great idea considering the way Intel's process leadership simply evaporated during the past few years.

I'm not holding my breath for 4.8Ghz @ TSMC but considering Zen is already doing 4.35Ghz stock on 12nm, a minimum 5% frequency jump is to be expected and a 10% jump is probable, unless you see Zen 3000 series to be all about low power 65W TDP SKUs. That 5-10% frequency jump means 4.5 - 4.8Ghz, or in other words we should reasonably expect the 3000 series to clock around 4.6-4.7Ghz.

 

[The Stilt]

Based on the software stack status, I find it extremely unlikely that any Zen 2 design besides Rome would currently be "pumping away" anywhere.

 

[DrMrLordX]

4.5ghz? so AMD is about to match frequencies Intel shipped 5 years ago.

Even worse, AMD can't even hit the same clocks as Piledriver. Su should just kill herself now, it's horrible.

Er wait.

No she shouldn't.

As far as Zen2's clockspeed limits/FMAX are concerned, I think we should bag on any predictions of clockspeeds in the range of 4.6-5.0 GHz. GF's 7nm process might have hit those clocks had GF completed their work. TSMC's process probably won't go past 4.5-4.6 GHz. It'll be good enough though, since relying on GF would have killed Matisse entirely.

On the plus side I think a modest AiO or even big air can probably get us 4.5 GHz all core with an overclock. In contrast, I think you can probably get 4.4 GHz all core with a 2700x or a 2950X with custom water . . .

 

[The Stilt]

As far as Zen2's clockspeed limits/FMAX are concerned, I think we should bag on any predictions of clockspeeds in the range of 4.6-5.0 GHz. GF's 7nm process might have hit those clocks had GF completed their work. TSMC's process probably won't go past 4.5-4.6 GHz. It'll be good enough though, since relying on GF would have killed Matisse entirely.

On the plus side I think a modest AiO or even big air can probably get us 4.5 GHz all core with an overclock. In contrast, I think you can probably get 4.4 GHz all core with a 2700x or a 2950X with custom water . . .

Any reason why you expect the GF 7nm LP would have been better than TSMC CLN7FF?
GlobalFoundries hasn't provided anything but failures since the 32nm SHP SOI got mature. What would have made them to succeed / meet the given promises this time around?

Personally I expect the TSMC process to be a great thing for Zen 2.
Not necessarily due to the (speculated) characteristics, but because the process provider isn't called GlobalFoundries.

Increasing the clocks from the current level we're are at (i.e > 4.7GHz on >= 8 cores) will be extremely hard, regardless if we are talking about Intel or AMD.
The smaller nodes will reduce the power dissipation, but they reduce the area as well. And we're already at a point where the intensity (power dissipation per area) is way too high, to allow the chips to be cooled efficiently.
Obviously it would be possible to improve the situation by using empty margins on the design, but I highly doubt that neither AMD or Intel is willing to spend expensive wafer space for the purpose.

 

[DrMrLordX]

Any reason why you expect the GF 7nm LP would have been better than TSMC CLN7FF?

POWER10. If I recall correctly - and maybe I don't - GF was planning on letting AMD fab on the same process as POWER10, which would have been an HP node. POWER chips tend to need/want high clockspeeds for the highest-end chips, which would have meant speeds at or near 5 GHZ. So the process limits would have to allow speeds that high. Everything else would have been design-related. AMD could have gotten there if they had wanted to.

GlobalFoundries hasn't provided anything but failures since the 32nm SHP SOI got mature. What would have made them to succeed / meet the given promises this time around?

Mostly true, but somehow they managed to deliver on HP nodes for POWER. Until now.

Interestingly enough, I've seen some recent announcements that POWER10 is now slated for 2020, on a 10nm process. Whose 10nm process is that? Originally it was supposed to be a 7nm process at GF. Obviously that isn't happening.

Personally I expect the TSMC process to be a great thing for Zen 2.

I'm not saying TSMC's 7nm process will be bad at all. I'm happy because it keeps Matisse on-schedule. GF was a potential delay threat, had GF continued development of their 7nm node and had AMD kept Matisse on that release schedule. The power characteristics also should be pretty good, and I think everyone will be impressed with the consistency of yields.

Not necessarily due to the (speculated) characteristics, but because the process provider isn't called GlobalFoundries.

Zing! I would like to see you in a boxing match with Hector Ruiz in an event similar to Uwe Boll vs Lowtax Kyanka. And no, I'm not just being a jerk - you'd probably curbstomp the old man. He would deserve it, too.

Of course I do not think he's really associated with GF anymore (except maybe as a shareholder).

Obviously it would be possible to improve the situation by using empty margins on the design, but I highly doubt that neither AMD or Intel is willing to spend expensive wafer space for the purpose.

It may be time for Intel and AMD to start using tiny heat pumps/TECs just to spread out heat at little. Or maybe IHSes that double as vapor chambers, or CNT IHSes, or . . . hell I don't know.

 

[maddie]

Increasing the clocks from the current level we're are at (i.e > 4.7GHz on >= 8 cores) will be extremely hard, regardless if we are talking about Intel or AMD.
The smaller nodes will reduce the power dissipation, but they reduce the area as well. And we're already at a point where the intensity (power dissipation per area) is way too high, to allow the chips to be cooled efficiently.
Obviously it would be possible to improve the situation by using empty margins on the design, but I highly doubt that neither AMD or Intel is willing to spend expensive wafer space for the purpose.

Actually I always thought that this is what Intel regularly did for the high performance models. Their actual transistor density was always, and still is, much lower than the theoretical quoted # for a given process.

AMD has also done something similar in Zen+, albeit for different reasons.

 

[french toast]

On a TSMC process? Has that ever happened with any of their processes? Even Global Foundries, who I believe were more focused on high performance nodes (Bulldozer, for example) couldn't keep up with Intel in the GHz race. No doubt, these upcoming chips will improve upon zen+ but something about TSMC and high 4.8GHz makes me real skeptical.

If that leak is true..and I am not even 70% sure that it is...then is it really hard to believe that AMD/TSMC could squeeze out another 200mhz F/max on single core using xfr2 over a buggy engineering sample?...I think not.

Also, even ignoring that leak. You think TSMC N7 HPC .. couldn't achieve a ~10% higher single core/dual turbo over a low power oriented '12'nm global foundries process?

Of course architecture is also going to play a part in that..but still you would be mightily disappointed if you were AMD and it couldn't go higher than 4.5ghz...

 

[Topweasel]

4.5ghz? so AMD is about to match frequencies Intel shipped 5 years ago. Intel is shipping 5ghz parts and it appears they could easily release 5.2-5.3ghz parts on 14nm. Intel's 10nm delays are horrible but, don't seem to be as bad now. It looks like they will remain competitive for the remainder of 2018/2019

OMG you are so right Intel released 8 core CPU's with 4.5 GHz stock clocks 5 years ago. How did I miss that? /s

7700K Max Turbo 4.5GHz.
6700K Max Turbo 4.2GHz.
2700X Max Turbo 4.3GHz.

7700k Release Date January 2017.