AMD to Begin Sampling 7nm "Zen 2" Processors Within 2018 for a 2019 Launch

[Captain_Tom]

Here we go again with this nonsense....You truly believe Ryzen 3 will bring as big of an upgrade as Ryzen from Excvator (55% performance uplift)? What planet do you come from? Your claims bear no realistic probability; it's not even close! You are claiming an improvement on an already existing architecture is supposed to bring larger performance improvements than Ryzen did from a 6-year-old architecture (which was already inferior for its time).

You claim that Ryzen will be "so far in front of any Intel offering it will be a painful mirroring of the xxxdozer vs Core days". Even very conservatively, we're talking about another 30%+ increase in IPC here, if it's gonna surpass Core in any way to actually make up your description. Meaning AMD will increase IPC by 30-40% and clock speed by 15% (again, both numbers are conservative).

So this slide by AMD



Is, clearly bollocks, huh? It's pretty evident that you, Captain_Tom, know better than AMD themselves, and AMD's graph showing around 15% overall performance uplift by Zen 2 is massive standing. Instead, the increase will be, as you said, like this:

Yeah...no.

1) So your response is to post an outdated roadmap with no specific markers on the Y-axis? LOL ok, great rebuttal! /s

2) Is English your first language? You seem to not understand what the word "could" means. I said Ryzen 3 could be as big as an uplift as Ryzen 1 was vs Excavator (~50% boost). I am only basing this "could" on recent rumors AND GloFo's own performance estimates:

https://www.extremetech.com/computing/263286-sitting-globalfoundries-talk-7nm-euv

^^^ The current estimate is that 7nm products can use 60% less power than 14nm for the same performance, or it can also offer a 40% boost at the same power consumption. There is also up to a massive 45% reduction in die size (almost half the size!). So even if we were to be insanely conservative and assume the end result is half as good as expected, we would get enough room to add 2-6 more cores and increase single-threaded performance by 40%.


3) So then let's get this straight - you think adding 50% more cores, and a massive IPC boost will not equal a 40-55% overall performance uplift. Well then I guess we can agree to disagree!


TLDR: It is not insane to think the R7 3800X could be a 12-core with 20% higher IPC and 4.8GHz clocks. That would destroy Intel's current plan of just slapping 2 more cores on Skylake again.

 

[HTC]

It is quite possible we could have 12 or 16 (I would guess 12 is more likely).

Actually one of the latest rumors is that AMD will separate the controllers, I/O, and core's onto their own separate dies; and then they will also make 3 and 4 core CCX's. The idea is that they will then have a 12nm I/O in the center, and then add 7nm 3 or 4-core CCX's as they wish around the other parts of the CPU.

TLDR - AMD is likely to have more than 8 cores with Ryzen 3, and they will likely make it even cheaper to produce than before by splitting up the CPU's into smaller components.

I've seen something similar in one of AdoredTV's videos (16 minute mark, for context):



Supposedly, the middle chip will be everything except the cores and the other 4 chips will be only the cores. Obviously, there should be infinity fabric in each chip to connect them.

Whether or not this is true or complete bollocks, i don't know ... but it is plausible. This should not increase core count, unless they also have CCXs in the corners, which isn't beyond the realm of possibilities: it could effectively double the core count.

 

[eidairaman1]

Could we be seeing 10 to 12 cores with Ryzen 2!?

A 12 core would need 2 cores disabled to get a 10core

 

[Deleted member 157276]

1) So your response is to post an outdated roadmap with no specific markers on the Y-axis? LOL ok, great rebuttal! /s

Outdated? It's literally 1 year old, and made by AMD themselves. And it perfectly represents reality, as their claim of the Zen+ performance increase in that "outdated" roadmap turned out to be right around what it was.

But if you claim it's "outdated" (apparently AMD managed to go from 15% performance increase to 55% performance increase in under a year), that means that there must be some newer evidence available, proving your claim of 50-55% performance increase. You are more than welcome to present that to us.

I am only basing this "could" on recent rumors AND GloFo's own performance estimates:

https://www.extremetech.com/computing/263286-sitting-globalfoundries-talk-7nm-euv

^^^ The current estimate is that 7nm products can use 60% less power than 14nm for the same performance, or it can also offer a 40% boost at the same power consumption.

GloFo also said the 12nm process would supposedly allow for "10% performance increase" with the same power delivery. In reality it barely provided half of thar. So we're talking about "up to 30%" from Ryzen 2, which again can be halved because of various limitations in place (mainly in architecture).

Process node shrinks that do theoretically allow for these performance increases almost never really ever do so in practice; you can look at Intel's different architectures over the past 6-7 years for proof of that, when they steadily went down from 32nm to 14nm. For example Ivy Bridge on 22nm did not provide any of the significant improvement in frequency over Sandy Bridge (same architecture) at 32nm, thought it should have been theoritically possible (in fact, IB had lower frequency). Why? Because you can't just arbitrarly provide theoritical improvements to actual ones. Same with Skylake at 14nm+ vs. Haswell at 22nm. By your logic, the base clock of the 6700K should have been at least 5.5 GHz base, when compared to the 2600K at 3.4 GHz, "because...uh...uh...32nm to 14nm provides over 70% possible performance improvement".

You can even compare GloFo's 7nm LP to TSMC or Samsung's 7nms, which according to specification we have gotten, is very similiar. TSMC's 7nm FinFET process will allow, according to themselves, "~20% speed improvement". How does it make sense that TSMC's 7nm process provides half the performance as GloFo?

Then there's the question about IPC, which you claim to be at least 20%. This claim alone matches/surpasses the performance increase in its entirety, as outlined by AMD, for Zen 2.

So even if we were to be insanely conservative and assume the end result is half as good as expected, we would get enough room to add 2-6 more cores and increase single-threaded performance by 40%.

Please tell how you came to this supendous conclusion, because I most certainly can't wrap my head around it.

1) S

TLDR: It is not insane to think the R7 3800X could be a 12-core with 20% higher IPC and 4.8GHz clocks.

20% IPC is an insane of an expectation. As is the idea of 12 full cores at 4.8 GHZ, when AMD can barely achieve 8 cores at 4 GHz on 12nm process with 115W TDP. 4 more cores is 50% more power usage; for the sake of the discussion, let's pretend the 7nm process will be able to offload all that through its better efficiency. 4.8 GHz on top of that is at least another 20% increase (although it certainly is way more, as power usage is not linear with frequency, as we all know) in power. Your estimates are therefore going beyond even GloFo's best-case numbers.

TLDR; everything you write is utter nonsense with no rational argument behind it, and is even debunked by AMD's own roadmap estimates (that so far have neither been updated, nor been proven to be wrong). You claim Zen 2 will perform almost 4x as much as AMD themselves claimed at a time when the Zen 2 was more or less complete.

 

[laszlo]

1) So your response is to post an outdated roadmap with no specific markers on the Y-axis? LOL ok, great rebuttal! /s

2) Is English your first language? You seem to not understand what the word "could" means. I said Ryzen 3 could be as big as an uplift as Ryzen 1 was vs Excavator (~50% boost). I am only basing this "could" on recent rumors AND GloFo's own performance estimates:

https://www.extremetech.com/computing/263286-sitting-globalfoundries-talk-7nm-euv

^^^ The current estimate is that 7nm products can use 60% less power than 14nm for the same performance, or it can also offer a 40% boost at the same power consumption. There is also up to a massive 45% reduction in die size (almost half the size!). So even if we were to be insanely conservative and assume the end result is half as good as expected, we would get enough room to add 2-6 more cores and increase single-threaded performance by 40%.


3) So then let's get this straight - you think adding 50% more cores, and a massive IPC boost will not equal a 40-55% overall performance uplift. Well then I guess we can agree to disagree!


TLDR: It is not insane to think the R7 3800X could be a 12-core with 20% higher IPC and 4.8GHz clocks. That would destroy Intel's current plan of just slapping 2 more cores on Skylake again.

i won't believe any rumors from GF as they promised a lot in the past and didn't delivered as expected.. of course maybe intel was in the same boat but we don't know all their shit

let's say GF "7nm" will behave like you wrote but you forget intel is also shrinking so the boost will roughly the same

i see different approach from each involved manufacturer (GF&Intel) and at this stage we don't know which one is better; difference can't be high as one has advantage over the other & viceversa

i wouldn't go deeper in this "black-hole" as we just speculate and "troll"

i'll quote something that catch my attention " Node names across the semiconductor industry have lost much of their meaning over the past 10 years " from here :https://www.eejournal.com/article/life-at-10nm-or-is-it-7nm-and-3nm/

peace and better working nodes!

 

[bug]

There may be nothing in production at this time, but I would argue that the future is, potentially, very bright for electronics. There are a number of exciting research projects in the works of which a sampling is here.

I know about (some of) those since I was in college. And that was a while ago. What we're missing is the transition to industrial scale production.

I even read somewhere that a research team managed to construct a transistor made of three atoms; unfortunately, I cannot find the reference ATM.

Well the Si atom is about .2nm in diameter. At 10 nm, that's 25 atoms side by side. Sure, the transistor is not a line, but you get the idea.

Btw, I believe your missing link is: https://www.theverge.com/2015/4/29/8515281/tmd-graphene-materials-science-ultrathin-electronics
To which I will add: https://en.wikipedia.org/wiki/Single-atom_transistor

But these are all proof of concepts. It's great we know they're possible, but we're realistically looking at at least a decade before we can put those to work. Provided we don't move off semiconductors by then.

Edit: For all those trying to show that the process from one manufacturer is better than the other's: they all source their lithography equipment from the same supplier (ASML), so there can't be significant differences there. You can tune it for better yields or better performance, but the output will still be in the same ballpark.

 

[sergionography]

Actually Intel 10 nm is almost same as Amd (GF) 7 nm....

but sound better 7 than 10 no?


View attachment 100487

Yes but 7nm glofo is 2 generations ahead of glofo 14nm. While intel 10nm is only one gen ahead of intel 14nm. So the progression is less. Also intel 10nm wont be moving to euv, unlike the glogo 7nm which is meant to transition to euv, so in other words; these glofo 7nm specs are only the beginning phase.

It's probably very safe to say that the GF 7nm process will be nowhere as good as Intel's 10nm, and I will put my life on the fact that we will not be seeing an 8+core Ryzen2 running air cooled at 5GHz anytime soon. For the next few years IPC is king for AMD, as Intel will match them on core count soon, and probably still better them on clock speeds. So AMD does need to aim and reach higher than the current Intel architecture, and settling for minor increments that result in deliberately staying below the IPC of what the current Intel architecture is producing is not what a company that very nearly went bankrupt a couple of years ago should be doing. This will only lead to a repeat of what AMD were doing when Intel released the Core2 architecture, which nearly destroyed AMD, and has taken them 10 years to counter. If they "settle" again, then AMD will not survive another 10 years of being significantly behind Intel. This time round, AMD no longer have a strong GPU division to back their financials up and provide turnover, unlike before as once again, after miss managing, they have lost to nVidia, and show no signs of being competent or able enough to do anything about it, AMD cannot afford to repeat past mistakes, it will end them.



This is critical for AMD, and just think about not taking into account Intel's next architecture! Intel are starting to wake up to AMD and the threat they once again pose. However they could crush AMD easily, and only by AMD aggressively pushing their architecture forward will AMD win over Intel.

14nm intel is a good half node or so ahead of everything else on 14nm in the market, perhaps even more. 10nm intel however wont be like that. 10nm intel will be slightly behind anything 7nm from other foundries, so that advantage wont exist anymore. Also i doubt 5ghz will be mainstream on anything as that's kind of the line we been stuck at since 32nm. Top potential clockspeed has not increased with shrinks, only lower consumption within that window (up to 5ghz). The only thing we will be seeing is high performance potential trickling down to mobile, having cores with bigger engines, having more cores, or having more robust socs with more specialty hardware.

1) So your response is to post an outdated roadmap with no specific markers on the Y-axis? LOL ok, great rebuttal! /s

2) Is English your first language? You seem to not understand what the word "could" means. I said Ryzen 3 could be as big as an uplift as Ryzen 1 was vs Excavator (~50% boost). I am only basing this "could" on recent rumors AND GloFo's own performance estimates:

https://www.extremetech.com/computing/263286-sitting-globalfoundries-talk-7nm-euv

^^^ The current estimate is that 7nm products can use 60% less power than 14nm for the same performance, or it can also offer a 40% boost at the same power consumption. There is also up to a massive 45% reduction in die size (almost half the size!). So even if we were to be insanely conservative and assume the end result is half as good as expected, we would get enough room to add 2-6 more cores and increase single-threaded performance by 40%.


3) So then let's get this straight - you think adding 50% more cores, and a massive IPC boost will not equal a 40-55% overall performance uplift. Well then I guess we can agree to disagree!


TLDR: It is not insane to think the R7 3800X could be a 12-core with 20% higher IPC and 4.8GHz clocks. That would destroy Intel's current plan of just slapping 2 more cores on Skylake again.

As beautiful as this sounds, i think its extremely difficult/impossible. Bringing over 20% ipc in one evolutionary update is simply not realistic. It was possible with first gen zen because bulldozer architecture simply tanked; having many obvious bottlenecks. Zen+ took about an extra 6-8 months (i assume if we believed the leaks) of work to simply extract 2-3% IPC, so over 20% basically needs a complete rework of the design or perhaps even a complete new architecture. I think however we will see 40-50% improvement in some workloads; such as avx for example, if amd adds 256 or 512 wide avx units instead of the current 128. Another area we could very likely see huge improvements is in multithreaded workloads. As it stands last i checked; amds smt implementation does better than intel, which is impressive considering the ccx units with the high latency are theoretically still negatively impacting performance. A reworked ccx model and infinity fabric could eliminate bottlenecks there which would make for some good returns.
Now to be clear, these things i mentioned are just some potential improvements that are currently possible according to my limited knowledge. Add to this better frequency scaling and u could have up to 40% in some workloads but more like 15-20% average at best.

 

[SkOrPn]

Zen 2 will be my second Ryzen chip (after Raven Ridge). I hope to push it all the way until the very last AM4 flagship which I hope will be Zen 3 on 7nm++, or even on 5nm perhaps (which I doubt will be any time soon). Then I will get the very last Zen 3 refresh iteration to replace the very first Zen 2 (the perfect longevity plan). I went with the STRIX X470-F Gaming because I have this feeling Zen 3 will introduce a super powerful "next-gen scalable" APU with built in HBM3 somewhere in the early 2020's, maybe a Renoir refresh but on 7nm++. If they do release a Ryzen 7 4600G/5600G (pure guessing here), or something along those lines I don't want to be caught with a mobo that does not have a Display Port on board. Lol

 

[T3cHn0pH1L3]

Outdated? It's literally 1 year old, and made by AMD themselves. And it perfectly represents reality, as their claim of the Zen+ performance increase in that "outdated" roadmap turned out to be right around what it was.

But if you claim it's "outdated" (apparently AMD managed to go from 15% performance increase to 55% performance increase in under a year), that means that there must be some newer evidence available, proving your claim of 50-55% performance increase. You are more than welcome to present that to us.



GloFo also said the 12nm process would supposedly allow for "10% performance increase" with the same power delivery. In reality it barely provided half of thar. So we're talking about "up to 30%" from Ryzen 2, which again can be halved because of various limitations in place (mainly in architecture).

Process node shrinks that do theoretically allow for these performance increases almost never really ever do so in practice; you can look at Intel's different architectures over the past 6-7 years for proof of that, when they steadily went down from 32nm to 14nm. For example Ivy Bridge on 22nm did not provide any of the significant improvement in frequency over Sandy Bridge (same architecture) at 32nm, thought it should have been theoritically possible (in fact, IB had lower frequency). Why? Because you can't just arbitrarly provide theoritical improvements to actual ones. Same with Skylake at 14nm+ vs. Haswell at 22nm. By your logic, the base clock of the 6700K should have been at least 5.5 GHz base, when compared to the 2600K at 3.4 GHz, "because...uh...uh...32nm to 14nm provides over 70% possible performance improvement".

You can even compare GloFo's 7nm LP to TSMC or Samsung's 7nms, which according to specification we have gotten, is very similiar. TSMC's 7nm FinFET process will allow, according to themselves, "~20% speed improvement". How does it make sense that TSMC's 7nm process provides half the performance as GloFo?

Then there's the question about IPC, which you claim to be at least 20%. This claim alone matches/surpasses the performance increase in its entirety, as outlined by AMD, for Zen 2.



Please tell how you came to this supendous conclusion, because I most certainly can't wrap my head around it.



20% IPC is an insane of an expectation. As is the idea of 12 full cores at 4.8 GHZ, when AMD can barely achieve 8 cores at 4 GHz on 12nm process with 115W TDP. 4 more cores is 50% more power usage; for the sake of the discussion, let's pretend the 7nm process will be able to offload all that through its better efficiency. 4.8 GHz on top of that is at least another 20% increase (although it certainly is way more, as power usage is not linear with frequency, as we all know) in power. Your estimates are therefore going beyond even GloFo's best-case numbers.

TLDR; everything you write is utter nonsense with no rational argument behind it, and is even debunked by AMD's own roadmap estimates (that so far have neither been updated, nor been proven to be wrong). You claim Zen 2 will perform almost 4x as much as AMD themselves claimed at a time when the Zen 2 was more or less complete.

Zen 2 (note Ryzen 2*** is Zen+ and not Zen 2 as some people are referring to it) was rumoured to have around 40% IPC increase compared to Zen(1) iirc from early AMD slides

 

[Deleted member 157276]

"[Zen 2 will give ] more like 15-20% average at best.

Buh...buh...muh 55% performance improvement! 12 cores at 4.8 GHz with 20% IPC increase!!!

Zen 2 (note Ryzen 2*** is Zen+ and not Zen 2 as some people are referring to it) was rumoured to have around 40% IPC increase compared to Zen(1) iirc from early AMD slides

Show me those slides, then we can talk. Otherwise, this simply isn't true, and the only AMD slides (which I have posted in this thread) show Zen 2 to increase performance (IPC + frequency) by around 15%.

 

[T3cHn0pH1L3]

Buh...buh...muh 55% performance improvement! 12 cores at 4.8 GHz with 20% IPC increase!!!



Show me those slides, then we can talk. Otherwise, this simply isn't true, and the only AMD slides (which I have posted in this thread) show Zen 2 to increase performance (IPC + frequency) by around 15%.

So a 15% improvement in IPC alone (dont confuse IPC with clock speed) is nothing to be sniffed at, heck, Ryzen1/+ is close to Intel's current IPC the only thing they lack behind is clock speed, if you had a Ryzen 1/+ at 5ghz then there would be no difference between AMD and Intel, just so happens that Ryzen is a new platform so is lacking in clock speed compared to Intel's 8th itteration, should they be able to maximise clocks there will be nothing in it, the only thing Intel has over AMD at the minute is those higher clock speeds so if we do see a 15% IPC improvement clock for clock over Ryzen + then clock speed becomes less of an issue, and if this is believed to be true as well as with a bumped up clock speed then AMD is set to equal or better current Intel's offerings. Still looking for said slide though it may well be a 40% IPC improvement over excavator from Ryzen and not Ryzen to Ryzen 2 as I first thought.

 

[Deleted member 157276]

So a 15% improvement in IPC alone (dont confuse IPC with clock speed) is nothing to be sniffed.

Where was IPC confused with clock speed? 15% IPC gains (which still is pretty unrealistic imo) was also not really the discussion, but rather 40% IPC claims. Or, in Captain_Tom's case, claims of 20% IPC + 50% more cores + 20-30% frequency increase (which he, I kid you not, claimed was "conservative").

The only time 15% was mentioned was in relation to performance, in this case AMD's own slides. That's IPC and frequency combined. I believe IPC increase of Zen 2 to be 6%+ and frequency to be 10%+, conservatively.

t, the only thing Intel has over AMD at the minute is those higher clock speeds so if we do see a 15% IPC improvement clock for clock over Ryzen + then clock speed becomes less of an issue,

Except even 15% IPC improvement is quite optimistic for a single generation and sizeable architecture improvement. If AMD is not able to achieve it, which it's most likey they won't, the other place to go is frequency. And why wouldn't they go with increased frequency; the transition to 7nm is literally only beneficial for adding more cores or/and increasing frequency within AMD's current 65-95W TDP constraints. AMD will in any case churn out as much performance as possible with Zen 2; if they got, say, 15% IPC improvement, that's not suddenly going to stop them from tapping into the possibility to increase frequency as well. AMD is breathing down Intel's neck with Zen+; surpassing Intel in SC performance will be huge achievment.

As mentioned, everything points towards the total performance improvement to be around 15% (hopefully a bit more), with IPC increase being in single digits. Any rational discussion about facts demands empirical data at the heart of any argument. That is, statements we make should be based on existing evidence. However much you like/dislike it, AMD's slides is the only and also closest description of Zen 2’s performance we currently have. And it specifically puts Zen 2 to give performance (IPC + clock speed) around 15%. Maybe AMD made significant breakthrough the last year and the end product will be 20%; maybe it will be 25%. We don’t know, and until we get more information, that 15% is the baseline.

40% IPC improvement over excavator from Ryzen and not Ryzen to Ryzen 2 as I first thought.

Correct. 40% was the initial proposed aim by AMD, when they first started making Zen (AMD was obviously not aiming at competing against Intel's latest and best architectures), which then grew to 52%. For Zen there's been no talk about IPC increase (maybe because AMD wants to surprise Intel?) But they did specify it as being complete early this year, with optimization and tuning being the only thing left.

 

[T3cHn0pH1L3]

Where was IPC confused with clock speed? 15% IPC gains (which still is pretty unrealistic imo) was also not really the discussion, but rather 40% IPC claims. Or, in Captain_Tom's case, claims of 20% IPC + 50% more cores + 20-30% frequency increase (which he, I kid you not, claimed was "conservative").

Well 8-10% on a minor node shrink 14nm to 12nm seems like 15% is very realistic in my book?

The only time 15% was mentioned was in relation to performance, in this case AMD's own slides. That's IPC and frequency combined. I believe IPC increase of Zen 2 to be 6%+ and frequency to be 10%+, conservatively.

Reviews and real world testing prove you wrong already, and this is just a node refresh, not even a revision...

Except even 15% IPC improvement is quite optimistic for a single (Intel) generation and sizeable architecture improvement. If AMD is not able to achieve it, which it's most likey they won't, the other place to go is frequency. And why wouldn't they go with increased frequency; the transition to 7nm is literally only beneficial for adding more cores or/and increasing frequency within AMD's current 65-95W TDP constraints. AMD will in any case churn out as much performance as possible with Zen 2; if they got, say, 15% IPC improvement, that's not suddenly going to stop them from tapping into the possibility to increase frequency as well. AMD is breathing down Intel's neck with Zen+; surpassing Intel in SC performance will be huge achievment.

10% came from a simple revision of Ryzen and a small node shrink, so another 15% is definitely on the cards with a new node, despite what you feel/think/anticipate.... Ryzen had a 4 year roadmap which has been right on track since the release of Ryzen 1*** why should it now change even if it doesnt live up to your expectations?

As mentioned, everything points towards the total performance improvement to be around 15% (where did you get 15% from? ) (hopefully a bit more), with IPC increase being in single digits. Any rational discussion about facts demands empirical data at the heart of any argument. That is, statements we make should be based on existing evidence. However much you like/dislike it, AMD's slides is the only and also closest description of Zen 2’s performance we currently have. And it specifically puts Zen 2 to give performance (IPC + clock speed) around 15%. Maybe AMD made significant breakthrough the last year and the end product will be 20%; maybe it will be 25%. We don’t know, and until we get more information, that 15% is the baseline.



Correct. 40% was the initial proposed aim by AMD, when they first started making Zen (AMD was obviously not aiming at competing against Intel's latest and best architectures) Wrong, AMD was spot on, even if it was comparing excavator against intel's 7th gen, it delivered that and more, surely you think 6c/12t would be here on Intel without Ryzen and you are a fool..) which then grew to 52% (It ended up 52% improvement over previous AMD CPU's) For Zen there's been no talk about IPC increase (maybe because AMD wants to surprise Intel?) But they did specify it as being complete early this year, with optimization and tuning being the only thing left.

Wrong again, IPC improvement was definitely mentioned after Ryzen+ both Ryzen and Ryzen+ have delievered exactly what they said they would, why would Ryzen2 be any different?

 

[Deleted member 157276]

Well 8-10% on a minor node shrink 14nm to 12nm seems like 15% is very realistic in my book?

You are mixing stuff here. Ryzen 2 didn't give us 8-10% IPC increase, it us 8-10% performance (IPC + frequency) increase. The actual IPC increase was 3%. So no, 15% IPC increase is not "very realistic", when we look at Ryzen 2. Not at all.

Reviews and real world testing prove you wrong already, and this is just a node refresh, not even a revision...

What are you on about? I'm talking about Zen 2, not Zen+ (Ryzen 2). How are Zen+ reviews proving anything I said about Zen 2 wrong? If you're referring to Zen+, the AMD slide notes around 10% performance increase for Zen +, which was what we got. A fact I've continously used as argument for the legitimacy and authority of the roadmap as a guideline for predictions.

10% came from a simple revision of Ryzen and a small node shrink, so another 15% is definitely on the cards with a new node, despite what you feel/think/anticipate.... Ryzen had a 4 year roadmap which has been right on track since the release of Ryzen 1*** why should it now change even if it doesnt live up to your expectations?

Yet again, you are mixing IPC with performance. The only one feeling and thinking is being done by you, my friend. My performance claims are grounded in AMD's roadmap. Your performance claims are grounded in...well feeling and thinking.

Ryzen had a 4 year roadmap which has been right on track since the release of Ryzen 1*** why should it now change even if it doesnt live up to your expectations?

What excactly have I said that contradicts the roadmap? You do realize that the roadmap claims 15% total performance (IPC + performance), and that it is rather you, not me, who is contradicting the roadmap?

Wrong again, IPC improvement was definitely mentioned after Ryzen+ both Ryzen and Ryzen+ have delievered exactly what they said they would, why would Ryzen2 be any different?

Since you clearly have zero understanding of the actual facts, let me remind you by posting the roadmap again:

Zen + provided almost 10% increase in performance (3% IPC + 6% frequency). Zen 2 is supposedly providing a bit better performance. My statement about 15% total performance (around 6% IPC and 10% frequency) is well in line with this roadmap -- It's actually somewhat higher, but I'm taking an extra year of tuning into account. I have from the very first moment based myself on this roadmap. Nothing else.

You are more than welcome to provide evidence where AMD ever claimed a 15% IPC improvement with Zen 2. I have a feeling you'll end up finding nothing, just as your initial 40% IPC claim...

 

[Captain_Tom]

I've seen something similar in one of AdoredTV's videos (16 minute mark, for context):

View attachment 100505

Supposedly, the middle chip will be everything except the cores and the other 4 chips will be only the cores. Obviously, there should be infinity fabric in each chip to connect them.

Whether or not this is true or complete bollocks, i don't know ... but it is plausible. This should not increase core count, unless they also have CCXs in the corners, which isn't beyond the realm of possibilities: it could effectively double the core count.

Yeah this is one of the big maybe's in the rumor pile regarding Ryzen 3 right now.

Another maybe to pay attention to is the rumor that AMD will make 2, 3, 4, and 6-core CCX variants. This would allow better cache coherency depending on the desired core count, and they could also make a 12-core desktop (or laptop!) chip. This would also allow interesting binning scenarios where they would make 4x2 = 8-Core chips with high clockspeeds and massive amounts of cache.

We will see

 

[CheapMeat]

Imagine if they put HBM or some other L4 system on the edges. This design reminds me a lot of what IBM does with their POWER chips.

 

[Captain_Tom]

View attachment 100517

Imagine if they put HBM or some other L4 system on the edges. This design reminds me a lot of what IBM does with their POWER chips.

Another one of those "maybe's"

 

[eidairaman1]

View attachment 100517

Imagine if they put HBM or some other L4 system on the edges. This design reminds me a lot of what IBM does with their POWER chips.

I say tighter packing, shorter traces...

 

[notanid]

Actually Intel 10 nm is almost same as Amd (GF) 7 nm....

but sound better 7 than 10 no?


View attachment 100487

Pitches are not relevant at all, total area of each transistor and yield density is the part that is relevant. Half that chart is 100% meaningless and looks like a misleading slide deck from intel.

 

[Captain_Tom]

Pitches are not relevant at all, total area of each transistor and yield density is the part that is relevant. Half that chart is 100% meaningless and looks like a misleading slide deck from intel.

Yep, and in fact the (third party) charts I have looked at suggested that Glofo's 7nm would roughly be equivalent to an "8.4nm" process from Intel.

Point is that the 7nm fabs AMD uses will be notably better than Intel's 10nm (even if Intel actually gets their 10nm to work lol). It will be comparable to the massive advantage Intel had over AMD in the 22nm vs 28nm days...

 

[sergionography]

Where was IPC confused with clock speed? 15% IPC gains (which still is pretty unrealistic imo) was also not really the discussion, but rather 40% IPC claims. Or, in Captain_Tom's case, claims of 20% IPC + 50% more cores + 20-30% frequency increase (which he, I kid you not, claimed was "conservative").

The only time 15% was mentioned was in relation to performance, in this case AMD's own slides. That's IPC and frequency combined. I believe IPC increase of Zen 2 to be 6%+ and frequency to be 10%+, conservatively.



Except even 15% IPC improvement is quite optimistic for a single generation and sizeable architecture improvement. If AMD is not able to achieve it, which it's most likey they won't, the other place to go is frequency. And why wouldn't they go with increased frequency; the transition to 7nm is literally only beneficial for adding more cores or/and increasing frequency within AMD's current 65-95W TDP constraints. AMD will in any case churn out as much performance as possible with Zen 2; if they got, say, 15% IPC improvement, that's not suddenly going to stop them from tapping into the possibility to increase frequency as well. AMD is breathing down Intel's neck with Zen+; surpassing Intel in SC performance will be huge achievment.

As mentioned, everything points towards the total performance improvement to be around 15% (hopefully a bit more), with IPC increase being in single digits. Any rational discussion about facts demands empirical data at the heart of any argument. That is, statements we make should be based on existing evidence. However much you like/dislike it, AMD's slides is the only and also closest description of Zen 2’s performance we currently have. And it specifically puts Zen 2 to give performance (IPC + clock speed) around 15%. Maybe AMD made significant breakthrough the last year and the end product will be 20%; maybe it will be 25%. We don’t know, and until we get more information, that 15% is the baseline.

To be fair, zen+ wasn't an architectural update, rather it was only a 12nm zen1 with some minor tweaks. Im not sure which website highlighted how amd didnt even reroute the layout for a smaller die and just practically left things as is with more empty space between transistors, indicating amd wasted no time here. So with this in mind, zen 2 basically is receiving 2 years(2 generations) worth of architectural work. Basically the work on zen+ being more inline of tweaking and testing new revisions or so; something that doesn't tie up senior engineers - allowing zen2 to be worked on in parallel. And with this being said, i still agree with you. Even with 2 years/generations worth of senior level engineering work - 15% ipc uplift is insanely difficult. Especially when you consider redesigning for 7nm(major shrink) and perhaps reconfiguring certain parts like adding more avx units or new instruction sets etc. Anywho ill say it again- i highly believe most of the improvements will in zen2 will be more related to the overall package than the cores themselves(better Infinity fabric and ccx design, better use of cache, higher throughput/lower latency memory controller, even more aggressive power management/turbo)

 

[Vya Domus]

It's probably very safe to say that the GF 7nm process will be nowhere as good as Intel's 10nm

Intel had so much trouble with this node for so much time , the only conclusion one can infer from that is that things are not going well for them at all.

https://www.semiwiki.com/forum/content/7191-iedm-2017-intel-versus-globalfoundries-leading-edge.html

They are fighting a losing battle with their manufacturing processes , sooner or later they'll will have to ditch their own fabs. These nodes have gotten so expensive and difficult to develop that not even someone like Intel can afford to dump so many resources for silicon that they don't sell to anyone else , unlike GloFo , TSMC , Samsung , etc.

 

[Deleted member 157276]

To be fair, zen+ wasn't an architectural update, rather it was only a 12nm zen1 with some minor tweaks. Im not sure which website highlighted how amd didnt even reroute the layout for a smaller die and just practically left things as is with more empty space between transistors, indicating amd wasted no time here. So with this in mind, zen 2 basically is receiving 2 years(2 generations) worth of architectural work. Basically the work on zen+ being more inline of tweaking and testing new revisions or so; something that doesn't tie up senior engineers - allowing zen2 to be worked on in parallel. And with this being said, i still agree with you. Even with 2 years/generations worth of senior level engineering work - 15% ipc uplift is insanely difficult. Especially when you consider redesigning for 7nm(major shrink) and perhaps reconfiguring certain parts like adding more avx units or new instruction sets etc. Anywho ill say it again- i highly believe most of the improvements will in zen2 will be more related to the overall package than the cores themselves(better Infinity fabric and ccx design, better use of cache, higher throughput/lower latency memory controller, even more aggressive power management/turbo)

AMD improved cache latency substantially, which led to the 3% increase of performance on same frequency over Ryzen 1 (this change was almost entirely responsible for the gaming performance increase of 10%). So IPC did in fact get better. The last 6% of performance were from clock speed improvements. You are however right; Zen+, just like Zen 3, is not an architectual improvement, but rather an optimization. Luckily for AMD, Zen is fresh enough for even "optimizations" of the same architecture providing larger performance uplifts than Intel has done between any generation in many years.

As for the time spent on Ryzen 2, it wasn't as much of "lazy" effort as people seem to think. The image being created is that AMD moved much of their focus from Zen+ over to Zen 2. That isn't entirely true. AMD have two different teams dedicated to their respective product lines. The main team that di Ryzen 1 is also doing Ryzen 3 (and possibly also 5), whereas the second team focuses on optimizations in Ryzen 2 and 4.

 

[ikeke]

https://www.semiwiki.com/forum/content/6895-standard-node-trend.html

Nice writeup on comparing processes between different semiconductor businesses.

 

[Vayra86]

1) So your response is to post an outdated roadmap with no specific markers on the Y-axis? LOL ok, great rebuttal! /s

2) Is English your first language? You seem to not understand what the word "could" means. I said Ryzen 3 could be as big as an uplift as Ryzen 1 was vs Excavator (~50% boost). I am only basing this "could" on recent rumors AND GloFo's own performance estimates:

https://www.extremetech.com/computing/263286-sitting-globalfoundries-talk-7nm-euv

^^^ The current estimate is that 7nm products can use 60% less power than 14nm for the same performance, or it can also offer a 40% boost at the same power consumption. There is also up to a massive 45% reduction in die size (almost half the size!). So even if we were to be insanely conservative and assume the end result is half as good as expected, we would get enough room to add 2-6 more cores and increase single-threaded performance by 40%.


3) So then let's get this straight - you think adding 50% more cores, and a massive IPC boost will not equal a 40-55% overall performance uplift. Well then I guess we can agree to disagree!


TLDR: It is not insane to think the R7 3800X could be a 12-core with 20% higher IPC and 4.8GHz clocks. That would destroy Intel's current plan of just slapping 2 more cores on Skylake again.

Really man? History proves you wrong, we're stuck at 4 Ghz mainstream and 5 Ghz OC for over a decade and several node shrinks; Ryzen's optimal power envelope abruptly ends at 4 Ghz too and the 2xxx series is only a baby step forward; it still has a significant TDP increase to achieve a marginally higher boost clock.

Core counts do not yield additional single thread performance but do increase the TDP, and clocking higher costs exponentially more energy. So the potential % gain in perf/watt only exists in each CPU designs' optimal power envelope, not outside it. Combine these facts and I honestly don't understand how you get to these bold statements, or whether you understand CPUs at all.

This 'could' happen in the same way flying unicorns 'could' happen.

View attachment 100517

Imagine if they put HBM or some other L4 system on the edges. This design reminds me a lot of what IBM does with their POWER chips.

And then there is this fairy tale of HBM stacks next to separated core dies... Sure, let's pop the HBM on precious pcb area instead of on an interposer the way its meant to be used for good measure Keep watching that AdoredTV nonsense... its great entertainment I get it.