GLOBALFOUNDRIES Introduces New 12nm FinFET Technology for High-Performance Applications

[itsmydamnation]

Yep. AMD will also get much more revenue from Pinnacle Ridge than it reasonably could from a Pinnacle Ridge-derived EPYC.

One of the new(unconfirmed) slides floating about had Zen2 EPYC at the end of 18 before the client release. So there isn't going to be much time for a PR server chip, if Zen2 was/is going to be later I would expect to see PR make its way to the server chips as it should add a nice amount of clock to all core base.

 

[Ajay]

One of the new(unconfirmed) slides floating about had Zen2 EPYC at the end of 18 before the client release. So there isn't going to be much time for a PR server chip, if Zen2 was/is going to be later I would expect to see PR make its way to the server chips as it should add a nice amount of clock to all core base.

Curious, one would think that the client would come first so that the best dice could be harvested (that plus more time to validate). In any case, if my bid to upgrade to Coffee Lake falls short (financially), I'm tempted to wait for Pinnacle Ridge, should be a nice CPU + have and upgraded chipset.

 

[itsmydamnation]

Curious, one would think that the client would come first so that the best dice could be harvested (that plus more time to validate). In any case, if my bid to upgrade to Coffee Lake falls short (financially), I'm tempted to wait for Pinnacle Ridge, should be a nice CPU + have and upgraded chipset.

The big issue for gen 1 EPYC was the time a complete ground up verification process takes . For gen 2 , things should be easier as so much of the work done for gen 1 gets to be the starting point for gen 2.

 

[Ajay]

The big issue for gen 1 EPYC was the time a complete ground up verification process takes . For gen 2 , things should be easier as so much of the work done for gen 1 gets to be the starting point for gen 2.

I would imagine AMD is going to do a new chipset, so there should be a fair amount of work. Still, AMD basically needed to rebuild the infrastructure vis-a-vis verifying their server components with various vendors - so that's all in place now.

 

[formulav8]

I'm expecting some IPC gains too:

I can see a BD to PD type IPC increase. 0%-7%. With some clockspeed with it. It could quite likely be just what the doctor ordered.

 

[Arachnotronic]

I can see a BD to PD type IPC increase. 0%-7%. With some clockspeed with it. It could quite likely be just what the doctor ordered.

There is no IPC increase in Pinnacle Ridge.

 

[formulav8]

There is no IPC increase in Pinnacle Ridge.

Not outright saying there is, but personal expectation, what makes you say there in't anything at all on ipc?

 

[Arachnotronic]

Not outright saying there is, but what is that based on?

AMD's own statement at their Financial Analyst Day. It's the same Zen core, even the same Summit Ridge architecture, just implemented in 12LP.

The improvements I would expect are in frequency/power consumption due to an improved physical layout and the superior transistor performance. Just like Kaby Lake.

 

[formulav8]

I think with the 15% higher advertised density, AMD will do a few tweaks, Whether IMC or other areas like buffers. At least it will be improved one way or another, even if only clocks like you expect.

 

[Ajay]

I think with the 15% higher advertised density, AMD will do a few tweaks, Whether IMC or other areas like buffers. At least it will be improved one way or another, even if only clocks like you expect.

I agree there will be tweaks, I'm guessing that the IMC will be better, but I doubt we'll see much in the way of IPC gains (not zero, just not enough to mention). May only be improvement in turbo core, overclocked memory and the like - no significant uarch changes.

 

[krumme]

I agree there will be tweaks, I'm guessing that the IMC will be better, but I doubt we'll see much in the way of IPC gains (not zero, just not enough to mention). May only be improvement in turbo core, overclocked memory and the like - no significant uarch changes.

Mike Clarke said back spring he had a good idea where the weak points is. To correct those they will have at most 6 months probably 2-3 from first 14nm tapeout results to implement before 12nm starts as 12nm seems a good deal different on paper specs. Very short time but there is easy gains in new arch. So basickly working 14nm 12 and 7 at the same time. I will expect some slight ipc uplift like avg 2%.
They have to stay within same tdp imo so its pretty limited what we can get stock but hopefully fmax is 10% up.
Or perhaps we will see a 120w tdp part or some more slack tdp ratings that Intel seems to think is the future for their brand. The latter posibilities is more marketing than engineering.
We need the products in hand.

 

[CatMerc]

There could still be a PPC improvement not from the core itself, but from an improved memory controller. A memory controller with lower latency would improve Ryzen's performance significantly in some applications.

 

[raghu78]

http://www.eetimes.com/document.asp?doc_id=1332328

Isn't this a big reduction?

14LPP > 12FDX = 47% power at ISO performance.

Ignoring area reduction, this 12 FDX process seems very close to 7LP in the performance parameters improvement over 14LPP

Yes 12FDX will be a very competitive and cost effective node for low power / low cost mobile and RF. btw the comparison above is of 14LPP 9T vs 12FDX 7.5T vs 7LP 6T . There is a 7LP 9T library option for high performance applications. So performance wise 7LP is a superior node vs 12FDX but at much higher complexity / cost.

 

[Ajay]

Mike Clarke said back spring he had a good idea where the weak points is. To correct those they will have at most 6 months probably 2-3 from first 14nm tapeout results to implement before 12nm starts as 12nm seems a good deal different on paper specs. Very short time but there is easy gains in new arch. So basickly working 14nm 12 and 7 at the same time. I will expect some slight ipc uplift like avg 2%.
They have to stay within same tdp imo so its pretty limited what we can get stock but hopefully fmax is 10% up.
Or perhaps we will see a 120w tdp part or some more slack tdp ratings that Intel seems to think is the future for their brand. The latter posibilities is more marketing than engineering.
We need the products in hand.

I sort of dismissed this at first, but after thinking about how GF and AMD sandbagged info on 12LP and Pinnacle ridge, I'm having second thoughts. We have no idea when the 12LP PDK made it into AMD's hands, but I suspect it is much earlier than we would think. The question, then , is how much time did AMD have to do more than a 'dumb' shrink. Again, going with the idea the AMD is sandbagging - I think they had a fair amount of time. Furthermore, I'm beginning to suspect that the '10%' number for a performance increase represents the floor of expectations (like the 40% number) and that the real result could be higher. I doubt it will earth shattering, but I wonder if the top end could be a 15% increase in performance. All in all, I now think that we can expect a very respectable jump in performance for Pinnacle Ridge. AMD is all in on Ryzen success and spending the money and manpower needed to stay at the top of their game and spoil Intel's lead. Interesting times indeed!

 

[coercitiv]

The question, then , is how much time did AMD have to do more than a 'dumb' shrink.

As already mentioned in the thread, all we need besides the obvious frequency bump is a perfected IMC: not only will that help unveil the true potential of the core, it will likely help it scale with frequency as well. Anything else is just icing on the cake.

 

[Ajay]

As already mentioned in the thread, all we need besides the obvious frequency bump is a perfected IMC: not only will that help unveil the true potential of the core, it will likely help it scale with frequency as well. Anything else is just icing on the cake.

I agree. I just happen t believe that 12LP and Pinnacle Ridge are going to be better than expected. Under promise and over deliver seems to be the CPU division's motto of late.

 

[scannall]

I agree. I just happen t believe that 12LP and Pinnacle Ridge are going to be better than expected. Under promise and over deliver seems to be the CPU divisions motto of late.

Makes sense though. No need to tip your hand for one. Also, if you came in even 1% under your promise the howls of outrage would be deafening.

 

[Ajay]

Makes sense though. No need to tip your hand for one. Also, if you came in even 1% under your promise the howls of outrage would be deafening.

Exactly. For AMD, coming out with anything less than a 10% performance increase will result in the outrage you mention - all the more reason to sandbag and do even better so no one can have any doubts about the performance increase and value of Ryzen on 12LP.

 

[IRobot23]

There is no IPC increase in Pinnacle Ridge.

Are you 100% sure about this?

 

[eddman]

I haven't read all the posts. What's the expected clock frequency improvement with 12nm?

 

[Abwx]

There is no IPC increase in Pinnacle Ridge.

That s not what Mark Papermaster once said, indeed the uarch is very versatile as they could gain some IPC "just" by improving the uops cache algorithms, so far Agner Fog point that the peak exe rate is 6 uops/cycle with the current CPU while HFR CPU reviewer (who has deep knowledge) said in an article that the theorical peak is 10 uops/cycle (8 for Haswell).

Papermaster confirmed the four-year lifespan and tapped the table in front of him: “We’re not going tick-tock,” he said. “Zen is going to be tock, tock, tock.”

https://www.pcworld.com/article/315...hitecture-is-expected-to-last-four-years.html

Surely that in January 2017 he already had a clear idea of what would be released on March 2018, it s not like CPUs tape outs and test silicon could be decided and released overnight.

 

[krumme]

I haven't read all the posts. What's the expected clock frequency improvement with 12nm?

GF claim 10% compared to other 14/16nm
"The new 12LP technology provides as much as a 15 percent improvement in circuit density and more than a 10 percent improvement in performance over 16/14nm FinFET solutions on the market today. "
If this is true and not the usual bragging that unfortunately is the standard its not the usual process rev but a more serious uplift. They are clearly also comparing to tsmc 16nm here so imo if they dont get that 10% vs 14lpp they run now its a disappointment imo.

 

[krumme]

That s not what Mark Papermaster once said, indeed the uarch is very versatile as they could gain some IPC "just" by improving the uops cache algorithms, so far Agner Fog point that the peak exe rate is 6 uops/cycle with the current CPU while HFR CPU reviewer (who has deep knowledge) said in an article that the theorical peak is 10 uops/cycle (8 for Haswell).



https://www.pcworld.com/article/315...hitecture-is-expected-to-last-four-years.html

Surely that in January 2017 he already had a clear idea of what would be released on March 2018, it s not like CPUs tape outs and test silicon could be decided and released overnight.

Yep its tock as a fact. And as Mike Clarke said he had a pretty good idea where it could be improved. (Perhaps the max uops)
But gaming perf is what most people understand as performance here so eg their own imc could be a major change.
20% gaming ipc uplift !!!!

 

[Arachnotronic]

That s not what Mark Papermaster once said, indeed the uarch is very versatile as they could gain some IPC "just" by improving the uops cache algorithms, so far Agner Fog point that the peak exe rate is 6 uops/cycle with the current CPU while HFR CPU reviewer (who has deep knowledge) said in an article that the theorical peak is 10 uops/cycle (8 for Haswell).



https://www.pcworld.com/article/315...hitecture-is-expected-to-last-four-years.html

Surely that in January 2017 he already had a clear idea of what would be released on March 2018, it s not like CPUs tape outs and test silicon could be decided and released overnight.

Same Zen core for Summit and Pinnacle. The only differences will be in frequency and possibly out of the box memory support.

 

[rainy]

Same Zen core for Summit and Pinnacle. The only differences will be in frequency and possibly out of the box memory support.

Since when marketing slides are indisputable proof?

Btw, do I need to remind you, that you have said Pinnacle Ridge will not show earlier than H2 2018 and at this moment Q1 is most probable?