Intel 10nm and GF 7nm at IEDM 2017

[Yotsugi]

while desktop Ryzen will use 7HPC.

That will make quite exquisite housefires.
AMD would never do that.
Besides, if they managed to get A72 running at 3.5, clocking wider, and deeper pipelined Zen would be silly easy.
HPC is an IBM-only node, designed to make POWER9+/POWER10 hit ~5GHz (current POWER9 barely managed 4GHz, and that's on FDSOI-based 14HP).

 

[Dayman1225]

IEDM 2017: Intel details 22FFL, a relaxed 14nm process for foundry customers, targets mobile and RF apps

 

[goldstone77]

Estimates
14nm 1.3 Normalized Frequency 7nm Soc 1.7 Normalized Frequency
1.3 to 1.7 is 30.7%@less power consumption
30.7% X 3.6GHz = 4.7052GHz
30.7% X 4.1GHz = 5.3587GHz
7nmHPC
1.7 to 2.1 is 23.5% @~2.5x power consumption
23.5% X 4.7052 = 5.810922GHz
23.5% X 5.3597GHz = 6.6179945GHz

Spoiler: normalized power/frequency graph

 

[CatMerc]

I think you could hear my maniacal laugh from the other side of the planet if AMD gets crazy enough to pull a 9590 done right and to release a 6GHz Zen CPU.

 

[Yotsugi]

I think you could hear my maniacal laugh from the other side of the planet if AMD gets crazy enough to pull a 9590 done right and to release a 6GHz Zen CPU.

Good.
The fire rises.

 

[raghu78]

Estimates
14nm 1.3 Normalized Frequency 7nm Soc 1.7 Normalized Frequency
1.3 to 1.7 is 30.7%@less power consumption
30.7% X 3.6GHz = 4.7052GHz
30.7% X 4.1GHz = 5.3587GHz
7nmHPC
1.7 to 2.1 is 23.5% @~2.5x power consumption
23.5% X 4.7052 = 5.810922GHz
23.5% X 5.3597GHz = 6.6179945GHz

Spoiler: normalized power/frequency graph

You should correlate GF's statements about GF 7LP low power being optimized for ARM A72 at 3.5 Ghz and 7HPC being designed for 5 Ghz operation.

https://www.globalfoundries.com/sites/default/files/product-briefs/7lp-product-brief.pdf

In that graph if you fit 3.5 Ghz for 1.4x the frequencies would fall in the range of realistically possible values.

1x = 2.5 Ghz
1.4x = 3.5 Ghz
2.1x = 5.25 Ghz

If we pick 1x = 2.4 Ghz 2.1x= 5.04 Ghz

The most efficient portion of 14LPP voltage freq curve was at 2.5-2.7 Ghz which is unsurprisingly the range where all core turbo for Ryzen mobile 2500u and the EPYC 7601 fall. In the graph 14LPP freq is plotted upto 1.3x or 3.3 Ghz which we know is the point beyond which 14LPP requires a lot of voltage for minimal clock gains.

https://www.notebookcheck.net/HP-En...00U-Radeon-Vega-8-Laptop-Review.266614.0.html

"Processor clock rate idles at 1.48 GHz when on Power Saver mode and can run up to 2.6 GHz and 3.6 GHz in multi-threaded and single-threaded operations, respectively, according to Task Manager."

https://www.anandtech.com/show/11544/intel-skylake-ep-vs-amd-epyc-7000-cpu-battle-of-the-decade/3

"The top part is the EPYC 7601, which is the CPU we were provided for in this comparison. This is a 32-core part with simultaneous multithreading, a TDP of 180W and a tray price of $4200. As the halo part, it also gets the good choice on frequencies: 2.20 GHz base, 3.2 GHz at max turbo (up to 12 cores active) and 2.70 GHz when all cores are active."

 

[Yotsugi]

You should correlate GF's statements about GF 7LP low power being optimized for ARM A72 at 3.5 Ghz and 7HPC being designed for 5 Ghz operation.

ARM core running at 3.5GHz on SoC node does not indicate any "optimization".
Heck, the same ARM core on 14LPP would run at ~2.3GHz, yet 14LPP peak efficiency for Zen is ~3.2GHz with 4.1GHz fmax.

 

[raghu78]

ARM core running at 3.5GHz on SoC node does not indicate any "optimization".
Heck, the same ARM core on 14LPP would run at ~2.3GHz, yet 14LPP peak efficiency for Zen is ~3.2GHz with 4.1GHz fmax.

Let me tell you that the peak efficiency for Zen core on 14LPP occurs in the 2.5 Ghz range. Thats exactly where the Ryzen mobile and EPYC parts turbo all their cores. btw please read the 7LP product brief . It clearly states high Performance (9T) for 5 Ghz operation and 6T > 17M gates/mm2 for maximum scaling. You must be out of your mind to think GF can get a high density H240 (6T) 2 fin cell to drive 5 Ghz clocks. Thats the job of H360 (9T) 4 fin cell.

https://www.globalfoundries.com/sites/default/files/product-briefs/7lp-product-brief.pdf

 

[Yotsugi]

Let me tell you that the peak efficiency for Zen core on 14LPP occurs in the 2.5 Ghz range.

I HEREBY SUMMON THE STILT AND HIS FMIN/FMAX CURVE.

Thats the job of H360 (9T) 4 fin cell.

The job of 7LP (HPC) is to make POWER9+/POWER10 hit 5GHz (it struggles hitting 4GHz on FDSOI-based 14HP).
That's it.
Got it?

 

[raghu78]

I HEREBY SUMMON THE STILT AND HIS FMIN/FMAX CURVE.

The job of 7LP (HPC) is to make POWER9+/POWER10 hit 5GHz (it struggles hitting 4GHz on FDSOI-based 14HP).
That's it.
Got it?

You refuse to even read GF's own manual and and keep repeating the same thing. I am done arguing with you. Nobody can get a high density 6T cell to drive 5 Ghz designs. Thats ridiculous. GF gave a comparison of Fmax of 14LPP 9T (CPP=78 MMP=64nm) vs 12FDX 7.5T (CPP=84nm MMP= 56nm) vs 7LP 6T (CPP=56nm MMP=40nm) at Vnom of the respective nodes. 1.0x vs 1.12x vs 1.06x.

https://m.eet.com/content/images/eetimes/1 7 12 14 copared x 800_1505972923.jpg

 

[Yotsugi]

Nobody can get a high density 6T cell to drive 5 Ghz designs.

"Nobody can get a high density 9T cell to drive 4GHz designs".

 

[raghu78]

"Nobody can get a high density 9T cell to drive 4GHz designs".

Zen on 14LPP HD is designed for 3.5 Ghz beyond which it takes an obscene amount of voltage for minimal clock gains before hitting a hard wall at 4.0Ghz. So first gen Ryzen is in no way a 4 Ghz design. Its a 3.5 Ghz design. the 14LPP product brief states > 3 Ghz which is quite true as we saw Stilt's Zen voltage/freq curve hit two points 3.3 and 3.5 Ghz where the voltage increase is non linear.

 

[goldstone77]

You should correlate GF's statements about GF 7LP low power being optimized for ARM A72 at 3.5 Ghz and 7HPC being designed for 5 Ghz operation.

https://www.globalfoundries.com/sites/default/files/product-briefs/7lp-product-brief.pdf

https://www.globalfoundries.com/sites/default/files/product-briefs/7lp-product-brief.pdf

It's a shame they don't have the 14LPP brief on the website to compare, which I believe said 3GHz operations. LP(lead-performance). I thought 7LP is based on GF own design, not based on Samsung 14LPP?

 

[Yotsugi]

I thought 7LP is based on GF own design

Yes.

 

[dfk7677]

From this:

and this:

I would say that the highest efficiency frequency is about 2.55GHz. But I am sure @The Stilt can enlighten us.

What we don't know is where is that on Total Power/Frequency graph. 0.65 normalized frequency or even less?

Edit:
In my opinion the 0.65 Normalized frequency is about 2.1GHz (and the highest point in the 14nm about 4.1GHz). So that gives us the highest frequency on 7nm SoC: 5.4GHz@95W power consumption.

 

[CatMerc]

Driving an A72 at 3.5GHz doesn't mean that it's the Fmax, just that it's an efficient point.

Some figures to take into account:
POWER 9 has a 12 stage pipeline
A72 has a 14 stage pipeline
Zen has a 19 stage pipeline

 

[el etro]

7HPC will make a nice Centurion(FX9000 series) processor. ~6.3Ghz 16 core at some 250W is genuinely nice.

 

[Yotsugi]

Some figures to take into account:
POWER 9 has a 12 stage pipeline
A72 has a 14 stage pipeline
Zen has a 19 stage pipeline

Ooh, someone finally started getting it.

 

[CatMerc]

Ooh, someone finally started getting it.

I "got it" long ago, but I do want to keep expectations at a reasonable level.
High clocking designs also require quite a bit of tuning to get right, as NVIDIA would tell RTG. Pipeline length and process performance mean nothing if you won't spend engineering team doing path optimization. And it's a /lot/ of time, and not really fun.

 

[CatMerc]

I remade this slide in order to reflect reality.

Slide removed due to it containing profanity.
Profanity (even in attachments) is not allowed
in tech forums.

After further thought, the slide is not likely
to be well received in an Intel thread, do not post
it again.

AT Mod Usandthem

 

[CatMerc]

This topic is both Intel 10nm and GF 7nm, and it is entirely relevant to the subject. It won't be well received only by those who don't want to see what the process comparison actually looks like. It is based on data from several sources coalesced into a more factual slide than Intel's marketing department provided.

Why would this be hidden?

I have made a slide replacing the problem word, so I can reupload it if given permission.

@UsandThem

 

[UsandThem]

This topic is both Intel 10nm and GF 7nm, and it is entirely relevant to the subject. It won't be well received only by those who don't want to see what the process comparison actually looks like. It is based on data from several sources coalesced into a more factual slide than Intel's marketing department provided.

Why would this be hidden?

@UsandThem

Mod decisions are not discussed in public threads.
If you wish to further discuss it, please create a post in
'Moderator Discussions'.

AT Mod Usandthem

 

[ksec]

GloFo's annual revenue is somewhere in the vicinity of $6 billion, Intel's is >$60 billion. The "GloFo has more R&D capability" meme is really silly because it's just not true.

Even TSMC's annual revenue is only in the neighborhood of $30 billion, or about half of Intel's. If Intel falls behind, it's not because GloFo/TSMC could outspend them, it's because Intel executed poorly.

Never did I implied GF has more R&D nor TSMC outspend Intel.
Intel has very different business model to both.
And I had always believe the revenue from GF doesn't truely show its full value due to agreement with AMD. Which we dont know in details and likely skew the numbers some what. Although insignificant when compared to Intel.
TSMC made more profit then Intel in the last quarter despite the *only* half of Intel's revenue.
Intel's poor execution is true,
But if the market wasn't booming with more transistors demand, TSMC and GF could never have catch up when Intel mis-step. That was the point I was making. There is a level of investment you need to reach before there is certain level of playing field. And this wasn't possible 10 years ago.

 

[ksec]

IEDM 2017: Intel details 22FFL, a relaxed 14nm process for foundry customers, targets mobile and RF apps

This looks ridiculously good. Any body did a comparison between this and 22FDX from GF?

 

[Dayman1225]

This looks ridiculously good. Any body did a comparison between this and 22FDX from GF?

I know GloFo did a direct comparison a while back but that was before Intel went from 7T to 6T so it is inaccurate. I don't know anyone who has recently done a direct comparison, but yeah it does look like quite a competitive and compelling node, we'll see if ICF actually gets some customers though.