New Zen microarchitecture details

[itsmydamnation]

AMD is a much smaller and has far fewer resources than Intel does. It's these unrealistic expectations of a frail, beaten-down company in the face of extremely strong, well-funded, and successful competition that leads to disappointment.

ahh i see , "coz AMD". Good to see you have continued to think through and expand your reasoning and logic.

 

[majord]

Hi guys, I am new to this forum, this is my first post here.

I've made some calculations based on your numbers and CB R15 scores, and they tell me that a 4c/4t 3400 MHz Zen would be equal to a 3100 MHz Sandy Bridge i5 (e.g. i5-2400).

If I take the Athlon X4 845 CPU's CB R15 score and give the 40% IPC boost it's still below but closer to the Sandy i5 performance.

A few months ago I said that if Zen reaches the Sandy's performance I would be glad with that. But is that all what we can expect from AMD's Zen?

Welcome!

Well firstly, there's absolutely no point focusing on one benchmark. 40% is an a average and the actual gain will vary between different workloads. Same applies right now comparing AMD to Intel. So, you have to look at as many relevant workloads as practical, and compute the average to have some idea what to expect.

That said, this is what I get using the 40% figure

CB15 - 1 core.

x4 845 @3.0Ghz : 78
Imaginary Zen @3.0Ghz: 109

From Anand's bench, using i3's since their clockspeed is fixed:

3MB Sandy @3.0Ghz: 107
(i3 2120)

3MB Haswell @2.9Ghz: 114
(i3 4130t)

i5's would be a tad faster due cache.

This differs from your results. Can you post more detail how you came to your figures?

If a new AMD Zen chip released in late 2016 / early 2017 can only hit the IPC speeds of an Intel chip released way back in 2011 I think that would be extremely disappointing. I think realistically they've got to hit Haswell 2013 / 2014 IPC to make it even competitive and then only with the usual price / core discounts relative to Intel.

given the fast number of other factors, I think basing your level on disappointment purely on IPC would be a bit silly.

 

[Doom2pro]

ahh i see , "coz AMD". Good to see you have continued to think through and expand your reasoning and logic.

Yeah these "cause reasons" excuses are getting old... AMD is known for getting more done with less... They are a fraction the size of Intel and nVidia and yet they aren't light years behind Intel and they are nipping at the heels of (and in some cases ahead) of nVidia.

I don't see why they can't beat Intel this round, or nVidia for that matter, they have done it before.

The only reason AMD looks to be dying is due to their stagnant CPU offerings the last 10 years, mainly due to a failed Bulldozer architecture (lesser with Phenom I issues) which they managed to make the best of with their semi-custom business (it could have been a LOT worse).

If they come out swinging with really competitive CPUs and GPUs this round, they are absolutely NOT dying and it's no surprise a lot of investors see AMD being $10 a share in 2017.

Imagine the semi-custom, HPC and Mobile (Laptop) business they are headed for if Zen and Polaris/Vega are the real deal... I'm salivating at the thoughts of a 4-8C/8-16T Zen based APU with Polaris/Vega GPU and HBM in a Laptop format.

 

[nenforcer]

Welcome!
given the fast number of other factors, I think basing your level on disappointment purely on IPC would be a bit silly.

Don't forget that Intel will have Z270 / Kaby Lake out later this year and by the time Zen hits late this year / early 2017 Intel Cannonlake .10nm will be right around the corner. The gap is only going to widen. AMD has a very narrow window of opportunity to get back in the game.

 

[deasd]

If a new AMD Zen chip released in late 2016 / early 2017 can only hit the IPC speeds of an Intel chip released way back in 2011 I think that would be extremely disappointing. I think realistically they've got to hit Haswell 2013 / 2014 IPC to make it even competitive and then only with the usual price / core discounts relative to Intel.

I can't see any significant difference between SanbyBridge and Skylake in 'IPC'.

 

[jhu]

I have a feeling we are going to have a long wait until the release. I'm guessing very late in 2016.First AMD has to show a big jump from Bulldozer/Piledriver.

But that's always been the case. Anyone who is not involved with making Zen can give nothing more than speculation. If Zen is coming 2H 2016, then we'll likely get performance leaks at that time.

 

[majord]

Don't forget that Intel will have Z270 / Kaby Lake out later this year and by the time Zen hits late this year / early 2017 Intel Cannonlake .10nm will be right around the corner. The gap is only going to widen. AMD has a very narrow window of opportunity to get back in the game.

Which will have zero (Kaby) to very little (Cannon) IPC uplift over Skylake.

Anyway that's irrelevant as my point was more that R&D budget aside, other factors such as physical size, Perf/watt, and clockspeed are all affected by increasing IPC.

Far fetched example, but would your disapointment be turned into excitment if IPC was 20% higher than skylake, but power consumption and die size meant they could only fit 6 of them on an SKU, and clock them at 2Ghz?

 

[frozentundra123456]

Yeah these "cause reasons" excuses are getting old... AMD is known for getting more done with less... They are a fraction the size of Intel and nVidia and yet they aren't light years behind Intel and they are nipping at the heels of (and in some cases ahead) of nVidia.

I don't see why they can't beat Intel this round, or nVidia for that matter, they have done it before.

The only reason AMD looks to be dying is due to their stagnant CPU offerings the last 10 years, mainly due to a failed Bulldozer architecture (lesser with Phenom I issues) which they managed to make the best of with their semi-custom business (it could have been a LOT worse).

If they come out swinging with really competitive CPUs and GPUs this round, they are absolutely NOT dying and it's no surprise a lot of investors see AMD being $10 a share in 2017.

Imagine the semi-custom, HPC and Mobile (Laptop) business they are headed for if Zen and Polaris/Vega are the real deal... I'm salivating at the thoughts of a 4-8C/8-16T Zen based APU with Polaris/Vega GPU and HBM in a Laptop format.

Ten dollars a share and they are going to beat both Intel and nVidia. OK then, sounds like a likely scenario.

 

[JDG1980]

Thanks for all the information. Your posts bring a welcome dose of reality to the sometimes wild speculation that seems to abound in these forums.

What The_Stilt posted about 14LPP was just as much "speculation" as everything that anyone else has said. He never claimed to have inside information, it was educated guesswork. So it seems like what you really object to isn't speculation per se, but speculation that paints AMD in a positive light.

 

[JDG1980]

As said before, these are my own expectations. I base them on the process characteristics (low power), process design targets (ASICs, mobile devices, network), on the fact how hard it is to get < 20nm processes right (look at Intel 14nm) and the various issues AMD has had with most of it's CPU designs in the past. Also if the low power process is the optimal choice for a relatively large and high power design such as Zen, why did Intel spend hundreds of millions in developing two separate process variants, one targeting power efficiency (low power) and the second one targeting high performance? Probably not just because they had nowhere else to put the incoming cash I recon.

We don't know what the actual efficiency differences are between Intel's 14nm process variants. Given how competitive the mobile space is, and its growing role in the computing marketplace, they could have felt it was worth the cash even if it only increased Core M's power efficiency by 5%. Or they could have estimated it would improve efficiency by 20% but it actually only did so by 10%. Keep in mind that TSMC was originally going to have a performance-focused 20nm process, but canned it when, upon testing, it turned out to have characteristics not much different from the SoC process.

And don't assume corporations are always rational with how they spend their money, especially if they're awash in money to be spent. Internal politics plays an important role as well.

Samsung is Korea's industrial titan, with annual profits twice as much as Intel's. I don't find it implausible that they could create a process that is competitive with Intel's best. (And Samsung definitely did create 14LPP; all GloFo has to do is not screw up the implementation too badly.)

 

[Arachnotronic]

Samsung is Korea's industrial titan, with annual profits twice as much as Intel's. I don't find it implausible that they could create a process that is competitive with Intel's best. (And Samsung definitely did create 14LPP; all GloFo has to do is not screw up the implementation too badly.)

How much do you think Samsung invests in logic R&D versus Intel?

 

[jhu]

Samsung is Korea's industrial titan, with annual profits twice as much as Intel's.

Samsung is a conglomerate. They make everything from ships to electric fans that automatically shutoff at night (because otherwise the fans will steal your soul, or something to that effect). Intel essentially only does chips.

 

[ElFenix]

Thanks for all the information. Your posts bring a welcome dose of reality to the sometimes wild speculation that seems to abound in these forums.

AMD is a much smaller and has far fewer resources than Intel does. It's these unrealistic expectations of a frail, beaten-down company in the face of extremely strong, well-funded, and successful competition that leads to disappointment.

ahh i see , "coz AMD". Good to see you have continued to think through and expand your reasoning and logic.

Yeah these "cause reasons" excuses are getting old... AMD is known for getting more done with less... They are a fraction the size of Intel and nVidia and yet they aren't light years behind Intel and they are nipping at the heels of (and in some cases ahead) of nVidia.

I don't see why they can't beat Intel this round, or nVidia for that matter, they have done it before.

The only reason AMD looks to be dying is due to their stagnant CPU offerings the last 10 years, mainly due to a failed Bulldozer architecture (lesser with Phenom I issues) which they managed to make the best of with their semi-custom business (it could have been a LOT worse).

If they come out swinging with really competitive CPUs and GPUs this round, they are absolutely NOT dying and it's no surprise a lot of investors see AMD being $10 a share in 2017.

Imagine the semi-custom, HPC and Mobile (Laptop) business they are headed for if Zen and Polaris/Vega are the real deal... I'm salivating at the thoughts of a 4-8C/8-16T Zen based APU with Polaris/Vega GPU and HBM in a Laptop format.

Ten dollars a share and they are going to beat both Intel and nVidia. OK then, sounds like a likely scenario.

What The_Stilt posted about 14LPP was just as much "speculation" as everything that anyone else has said. He never claimed to have inside information, it was educated guesswork. So it seems like what you really object to isn't speculation per se, but speculation that paints AMD in a positive light.

How much do you think Samsung invests in logic R&D versus Intel?

can you guys quit ruining this thread? we had a nice couple days where some semi-informed speculation went on, but you all had to go and step in it. you've had this discussion over and over in thread after thread, none of you are changing any of the others minds, and i'm tired of reading it.

 

[Dresdenboy]

It doesnt really matter what was the test vehicle as they didnt release 14nm numbers in isolation but in respect of both their 28nm HPP and SLP, from the numbers above we can extract directly the reduction in parasistic capacitance and the improvement in conductance, indeed i just checked in a recent slide where they summarized thoses numbers in more accessible way, they claim in respect of 28 HPP :

1.521x speed at isopower.
0.442x power at isospeed.
0.495x density reduction.

Set apart density the other numbers are directly extracted from the ones they published in their engineers dedicated conf, they specify that the 14nm LPP is well suited for high performance designs.

Is the density factor for a single dimension or the area?

Here is the GF presentation: https://www.cadence.com/cadence/eve...ing_Signoff_Richard_Trihy_Globalfoundries.pdf

 

[AtenRa]

Don't forget that Intel will have Z270 / Kaby Lake out later this year and by the time Zen hits late this year / early 2017 Intel Cannonlake .10nm will be right around the corner. The gap is only going to widen. AMD has a very narrow window of opportunity to get back in the game.

Cannonlake will launch in Mobile first in Q2-Q3 2017. Since Kabylake is Q4 2016 dont expect a 10nm Desktop before Q4 2017 or even Q1 2018.

 

[guskline]

Per the Super Moderator's suggestion, and not being in the "inside" of testing my question to Dresdenboy and others at this time is as follows concerning the Zen Architecture.

Assume they have produced some ES sample chips, where would they ship them to be tested? Would AMD supply the MBs since it is a new socket? Is there a specific protocol to testing a cpu and does that change with each new class of chips?

I've been lucky enough to be used for testing a few pieces of software and there are lots of reports to be filed. I suspect testing of a new cpu such as this is much more intense.

 

[Abwx]

Is the density factor for a single dimension or the area?

Here is the GF presentation: https://www.cadence.com/cadence/eve...ing_Signoff_Richard_Trihy_Globalfoundries.pdf

That s a 2013 doc, the one i refered to was published in late 2015 and is a comparison between 14nm LPP, 28nm HPP/SLP/LPS and 20nm LPM, that s a summary of the conf that can be found on Youtube.

They quoted the area of 20nm LPP as 0.50x/0.55x the one of the 28nm LPS/SLP-HPP and 14nm being actualy 0.55x/0.60x in respect of said 28nms, there s possibly a typo in their slide because they report 10% higher density for 20nm LPM than for 14nm LPP but it s also quite possible.

 

[Dresdenboy]

That s a 2013 doc, the one i refered to was published in late 2015 and is a comparison between 14nm LPP, 28nm HPP/SLP/LPS and 20nm LPM, that s a summary of the conf that can be found on Youtube.

I know, I posted the slides from that video here. I didn't see it on YT, but something else, Vimeo maybe.

 

[Yutani]

Welcome!

Well firstly, there's absolutely no point focusing on one benchmark. 40% is an a average and the actual gain will vary between different workloads. Same applies right now comparing AMD to Intel. So, you have to look at as many relevant workloads as practical, and compute the average to have some idea what to expect.

That said, this is what I get using the 40% figure

CB15 - 1 core.

x4 845 @3.0Ghz : 78
Imaginary Zen @3.0Ghz: 109

From Anand's bench, using i3's since their clockspeed is fixed:

3MB Sandy @3.0Ghz: 107
(i3 2120)

3MB Haswell @2.9Ghz: 114
(i3 4130t)

i5's would be a tad faster due cache.

This differs from your results. Can you post more detail how you came to your figures?

given the fast number of other factors, I think basing your level on disappointment purely on IPC would be a bit silly.

As I observed the scores of different uarchs I've found a pattern and based on it I was able to create an Excel chart where (in a limited form) I can calculate the single and multi thread performances of each uarch. It's not valid for SMT-capable CPUs and I can not do anything with turbo frequency results because they distort the numbers. This model can be applied on real CPU cores and AMD modules (CUs), and can handle fixed clock ST and MT results. So it's not a big deal.

I made two scenarios. First was when 2 Zen cores outperform an Excavator CU by 40%. This was resulting a speed right behind Sandy Bridge.

The second scenario was when I added 40% to the Athlon X4 845's ST performance. (And now I can see that I've made a mistake here.) In this case Zen's performance would be between Ivy Bridge and Haswell.

 

[Phynaz]

Is the density factor for a single dimension or the area?

Here is the GF presentation: https://www.cadence.com/cadence/eve...ing_Signoff_Richard_Trihy_Globalfoundries.pdf

Hahaha...GloFo said they would be on 10nm last year.

Sure makes their marketing believable.

 

[Dresdenboy]

Hahaha...GloFo said they would be on 10nm last year.

Sure makes their marketing believable.

Planning the unknown. This is part of the business, as there's always the other side asking "When?".

 

[Abwx]

I know, I posted the slides from that video here. I didn't see it on YT, but something else, Vimeo maybe.

Dont know Vimeo but on YT they state 72% area reduction from 28 HPP to 14nm LPP, at 37min :

https://www.youtube.com/watch?v=wwa_GqU2LlQ

So 50-55% better density with complex chips is quite possible.

 

[Dresdenboy]

Dont know Vimeo but on YT they state 72% area reduction from 28 HPP to 14nm LPP, at 37min :

https://www.youtube.com/watch?v=wwa_GqU2LlQ

So 50-55% better density with complex chips is quite possible.

Yep, that's the same video. Good to see it on YT, so I can D/L it. ^^

I noticed, that they used 9T libs for the 28HPP instance. This means, with 12T the block would have delivered different results. Also the total power of the 28HPP instance is the same as of the 14LPP instance, the latter just with a much better leakage and frequency.

 

[Dresdenboy]

As I observed the scores of different uarchs I've found a pattern and based on it I was able to create an Excel chart where (in a limited form) I can calculate the single and multi thread performances of each uarch. It's not valid for SMT-capable CPUs and I can not do anything with turbo frequency results because they distort the numbers. This model can be applied on real CPU cores and AMD modules (CUs), and can handle fixed clock ST and MT results. So it's not a big deal.

I made two scenarios. First was when 2 Zen cores outperform an Excavator CU by 40%. This was resulting a speed right behind Sandy Bridge.

The second scenario was when I added 40% to the Athlon X4 845's ST performance. (And now I can see that I've made a mistake here.) In this case Zen's performance would be between Ivy Bridge and Haswell.

Yep, the module's scaling (in CB MT a hybrid of a 2C CMP and a SMT FPU) distorts the picture.

Have you seen the article by Looncraz?
http://excavator.looncraz.net/
While seeing Looncraz's remark about Zen's possible integer performance vs. SNB, I think, that it might even be higher, so why not? FP - especially with 256b AVX stuff - will "suffer" (i.e. have lower IPC than HSW -> ignoring that with wider SIMD, less instructions are being executed per cycle).

 

[Abwx]

Yep, that's the same video. Good to see it on YT, so I can D/L it. ^^

I noticed, that they used 9T libs for the 28HPP instance. This means, with 12T the block would have delivered different results. Also the total power of the 28HPP instance is the same as of the 14LPP instance, the latter just with a much better leakage and frequency.

They had to use nearby power to have the same temps and relative perfs degradation for both processes, otherwise getting the 28 HPP at same frequency than 14 LPP would increase power by (2.41/1.17)^2 = 4.23x.

The dynamic power would be 888mW (with leakage increasing by 50%..), that is 2.86x more than 14 LPP.

If we add leakage in the balance the power would be roughly 1W+, at this power the 14nm LPP would clock at 4.3GHz in the same experiment...