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

[Ajay]

I doubt it. R7s have already shown the ability to run IF speeds up to 2000 MHz, stable. The problem is the IMC and/or RAM ICs can't handle DDR4-4000 at the present. There may be some bizarre technical reason why they are unable to choose different ratios, but in terms of raw speed, the present IF implementation CAN handle such speeds. Plus there's no reason why they couldn't fix that technical limitation, if they so chose.

Sorry, I should have said DF/DRAM speed ratio.

 

[IRobot23]

I doubt it. R7s have already shown the ability to run IF speeds up to 2000 MHz, stable. The problem is the IMC and/or RAM ICs can't handle DDR4-4000 at the present. There may be some bizarre technical reason why they are unable to choose different ratios, but in terms of raw speed, the present IF implementation CAN handle such speeds. Plus there's no reason why they couldn't fix that technical limitation, if they so chose.

Basically if they could optimized it for 2:1 IMC speed it would be glorious clock up top 3600MHz would be great... well dreaming 🙄

 

[raghu78]

Sorry, I should have said DF/DRAM speed ratio.

We don't know yet if there are any technical limitations in the Zen architecture which prohibit a higher DF DRAM ratio or if the Zen architecture can even support a different and fixed fabric speed. Pinnacle Ridge will give us a clue as to how flexible the Zen architecture is. If AMD can change the DF / DRAM speed ratio from 1:1 to 2:1 or 3:2 then it would be a huge improvement to latency. I think the sweet spot for memory latency and power is likely to be 2.4 Ghz DF speed. Improving memory support to DDR4 4000 is also something which we can look forward to.

 

[raghu78]

http://www.digitimes.com/news/a20171024PD215.html

"Globalfoundries is also planning the rollout of its 12nm leading-performance (12LP) FinFET manufacturing process to deliver better density and a performance boost over its current-generation 14nm FinFET offering. The 12LP node will be ready for volume production in the first quarter of 2018, according to the foundry."

It looks like Pinnacle Ridge should arrive in late Q1 or early Q2 2018.

 

[PeterScott]

http://www.digitimes.com/news/a20171024PD215.html

It looks like Pinnacle Ridge should arrive in late Q1 or early Q2 2018.

Digitimes trades in rumors not facts.

What GF actually says, is that 12LP is in "Risk Production" for 1H 2018. So I wouldn't expect quick transition to 12nm early in 1H.

 

[Yotsugi]

What GF actually says, is that 12LP is in "Risk Production" for 1H 2018.

And the other (too lazy to find it) GF slide says volume H1.

 

[PeterScott]

And the other (too lazy to find it) GF slide says volume H1.

If it does, it's probably an older slide. Fabs almost always change estimates in one direction: Later.

 

[Yotsugi]

If it does, it's probably an older slide. Fabs almost always change estimates in one direction: Later.

You don't delay something that is a glorified node tweak.
Stop being so pessimistic.

 

[Mulrian]

A lot of AMD fans are expecting at least a 10% improvement in turbo clocks over current Ryzen with the new 12nm process. That would be great, but I think quite wishful thinking to be honest.

 

[PeterScott]

You don't delay something that is a glorified node tweak.
Stop being so pessimistic.

It isn't pessimistic to take cues from a recent (Sept 2017), official slide, from Global Foundries.

I'll believe something to the contrary, when GF says something to the contrary.

 

[scannall]

A lot of AMD fans are expecting at least a 10% improvement in turbo clocks over current Ryzen with the new 12nm process. That would be great, but I think quite wishful thinking to be honest.

While I don't think that a %10 uplift in turbo clocks is unreasonable, like all other speculation I read it for amusement more than anything. I keep my expectations low, so I can be pleasantly surprised if things come out better. At any rate, we'll know for sure in a few months.

 

[Yotsugi]

A lot of AMD fans are expecting at least a 10% improvement in turbo clocks over current Ryzen with the new 12nm process. That would be great, but I think quite wishful thinking to be honest.

Official node characteristic is a wishful thinking?
The fuck.

 

[Mulrian]

They are comparing against 'industry 16nm FinFET solutions' so could be anything. Do you really think we will see Ryzens going up to 4.4ghz with this node change?

 

[Ajay]

They are comparing against 'industry 16nm FinFET solutions' so could be anything. Do you really think we will see Ryzens going up to 4.4ghz with this node change?

Don't know, but I'm pretty sure AMD will get at least 10% more throughput out of Ryzen, one way or another (I actually expect more, as I think they are sandbagging). AMD needs to have a response to Coffee Lake and Ryzen on 12LP is it. We'll see soon enough.

 

[The Stilt]

GlobalFoundries states that compared to 14nm LPP, the 12nm LP has 10% higher performance at ISO power.
Better hope that we see increased Fmax with it, instead of just lower power consumption at. e.g. 3GHz 🙄

 

[DrMrLordX]

Pinnacle Ridge won't find purchase in any of AMD's server products, so I expect that it'll be tuned with the workstation/DiY market in mind. It would be strange if all they did was re-release Summit Ridge with a better voltage curve at 75% of Fmax.

 

[Ajay]

GlobalFoundries states that compared to 14nm LPP, the 12nm LP has 10% higher performance at ISO power.
Better hope that we see increased Fmax with it, instead of just lower power consumption at. e.g. 3GHz 🙄

There may be some small tweaks that improve performance as well - would make sense while re-synthesizing the die for 12lp.

 

[CatMerc]

If it does, it's probably an older slide. Fabs almost always change estimates in one direction: Later.

It's not older, it was from the same day.
Just GloFo being confusing.

I theorized it may be that due to AMD working closely with GloFo, they can enter volume production rather early, while any other partner will only get risk production at the same time.

Pinnacle Ridge won't find purchase in any of AMD's server products, so I expect that it'll be tuned with the workstation/DiY market in mind. It would be strange if all they did was re-release Summit Ridge with a better voltage curve at 75% of Fmax.

Would be interesting to see Pinnacle with half of the GMI links cut (one is left for Threadripper), and any other server feature in those dies, as well as optimizing for Fmax.

 

[Spartak]

Look at this crap from thg
"However, this is nothing really new, as Intel has been ahead in process technology for quite some time, and if anything, third-party foundry companies such as Samsung/Global Foundries and TSMC have closed the gap a little once they adopted the FinFET structure for their transistors. "
Do they take us for morrons. Thg is nothing but a whore these days. But what should we complain as customers that dont even pay.

That ticks me as well. But, let them be. As long as they able to keep their website free to the readers, I don't mind if they do "sponsored" news occasionally.

Could one of you two explain me what is exactly wrong about that quote?

1. Globalfoundries just introduced FinFET this year with Ryzen, Intel uses it already since 2012.
2. Intel had a process lead since, literally day zero.
3. There is still a process node performance gap between Intel and GF as the 14nm process used by Intel has a much smaller MPxGP

This comes from someone that a. stopped reading THG back in '98 and b. is still using a Phenom II waiting for Raven Ridge. But boy, did I just land in some fanboy process node alternate reality?

The only possible word to take offence in is the word "little". If that's the case that's just sad.

 

[Phynaz]

But boy, did I just land in some fanboy process node alternate reality?

Yes.

 

[SpaceBeer]

...
1. Globalfoundries just introduced FinFET this year with Ryzen, Intel uses it already since 2012.
....

1. GloFo introduced mass FinFET production last year with Polaris
2. AMD just introduced SMT this year with Ryzen, Intel uses it already since 2002, i.e. for 15 years. AMD's implementation is more or less the same. In some workloads HT works better, in some AMD's SMT has advantage.

I'm not saying intel's process isn't better, cause it is. Just that this "time argument" tells nothing about the difference.

 

[Spartak]

1. GloFo introduced mass FinFET production last year with Polaris
2. AMD just introduced SMT this year with Ryzen, Intel uses it already since 2002, i.e. for 15 years. AMD's implementation is more or less the same. In some workloads HT works better, in some AMD's SMT has advantage.

I'm not saying intel's process isn't better, cause it is. Just that this "time argument" tells nothing about the difference.

The time difference was mentioned in relation to the quote "Intel has been ahead in process technology for quite some time"

1. The THG context was about AMD processors, not GPUs. But even then, it'd be 4 years instead of 5. Your point in disproving the original claim was....?
2. SMT has literally nothing to do with process technology. SMH.

Please, in your next try, on another topic, try to answer the original question without introducing irrelevant tangents.

 

[ajc9988]

1. GloFo introduced mass FinFET production last year with Polaris
2. AMD just introduced SMT this year with Ryzen, Intel uses it already since 2002, i.e. for 15 years. AMD's implementation is more or less the same. In some workloads HT works better, in some AMD's SMT has advantage.

I'm not saying intel's process isn't better, cause it is. Just that this "time argument" tells nothing about the difference.

Could one of you two explain me what is exactly wrong about that quote?

1. Globalfoundries just introduced FinFET this year with Ryzen, Intel uses it already since 2012.
2. Intel had a process lead since, literally day zero.
3. There is still a process node performance gap between Intel and GF as the 14nm process used by Intel has a much smaller MPxGP

This comes from someone that a. stopped reading THG back in '98 and b. is still using a Phenom II waiting for Raven Ridge. But boy, did I just land in some fanboy process node alternate reality?

The only possible word to take offence in is the word "little". If that's the case that's just sad.

Actually, finFET was under patent exclusion, which Intel did not use the patent for the majority of the patent period. Then, toward the end, they used it. After Intel used it, others looked for ways to do it without tripping on the patent, but the majority waited out the clock.

Not talking about process, on threading, SMT actually beats HT. Period. Where Intel still outperforms is IPC and speed. This is why Ryzen seems to catch up in heavily multithreaded benches in some instances, even though running slower. It is whether the multithreaded capabilities approach overcoming other advantages that Intel has. It is very easy to conflate the different capabilities and aspects of chips! ;-)

 

[krumme]

Could one of you two explain me what is exactly wrong about that quote?

1. Globalfoundries just introduced FinFET this year with Ryzen, Intel uses it already since 2012.
2. Intel had a process lead since, literally day zero.
3. There is still a process node performance gap between Intel and GF as the 14nm process used by Intel has a much smaller MPxGP

This comes from someone that a. stopped reading THG back in '98 and b. is still using a Phenom II waiting for Raven Ridge. But boy, did I just land in some fanboy process node alternate reality?

The only possible word to take offence in is the word "little". If that's the case that's just sad.

Samsung have been producing 10nm for the most of the year.
And tsmc 10nm for iphone for next to half a year.
You forgot that? I think so 😉 You were looking at x86 but the volume today is elsewhere.
7nm that is more equivalent to Intel 10nm is in booting up in production now. 7nm euv from samsung is comming in a good year.
Tsmc 7nm euv and 5nm is probably going to be the stars in 2020 forward.

The days when Intel was in front is gone.

Heck even the days when an Intel engineer visited Anand to bust the x86 power myth is gone.

Now Anand is working at apple and tsmc is well into 5nm with euv.

Better wake up to the new world.

 

[PeterScott]

Not talking about process, on threading, SMT actually beats HT. Period. Where Intel still outperforms is IPC and speed. This is why Ryzen seems to catch up in heavily multithreaded benches in some instances, even though running slower. It is whether the multithreaded capabilities approach overcoming other advantages that Intel has. It is very easy to conflate the different capabilities and aspects of chips! ;-)

The only reason Ryzen catches up while running slower is because it is 8 cores vs 6.

Show me even one example where a Ryzen 6C/12T catches up with Intel CL 6C/12T.