Question How ahead is Intel in CPU design compared to AMD?

[Adonisds]

If we could get both companies to design a CPU core now on the same process and using the same number of transistors, who would be ahead and by how much? I'm assuming Intel would be ahead because Skylake has a similar IPC to Zen 2, but Zen 2 uses a more dense node, and Skylake is from 2015 while Zen 2 is from 2019.

The both companies using the same process scenario is just to illustrate the problem and I know it will never happen, but maybe there is a smart way to compare current CPU designs and find out who is ahead and by how much. Wikichip has die sizes and die pictures of Sunny Cove and Zen 2, but I'm not sure how to interpret and compare them.

Another question: If Intel had to design future cores using the same process and the same die size as Skylake, how much further could the Skylake design be improved? Also assuming it will have to have to be x86 too and have all the same functions.

 

[Markfw]

If we could get both companies to design a CPU core now on the same process and using the same number of transistors, who would be ahead and by how much? I'm assuming Intel would be ahead because Skylake has a similar IPC to Zen 2, but Zen 2 uses a more dense node, and Skylake is from 2015 while Zen 2 is from 2019.

The both companies using the same process scenario is just to illustrate the problem and I know it will never happen, but maybe there is a smart way to compare current CPU designs and find out who is ahead and by how much. Wikichip has die sizes and die pictures of Sunny Cove and Zen 2, but I'm not sure how to interpret and compare them.

Another question: If Intel had to design future cores using the same process and the same die size as Skylake, how much further could the Skylake design be improved? Also assuming it will have to have to be x86 too and have all the same functions.

This has nothing to do with reality at the moment, so why ask this ?

 

[Adonisds]

This has nothing to do with reality at the moment, so why ask this ?

Curiosity, fun, I could learning something about CPU design.

Do you often visit question threads just to tell the OP that you don't like the question?

 

[dmens]

They are not. AMD has clearer vision, up-to-date methodology and smarter people.

 

[JasonLD]

I am sure Intel has better products planned for the future, however, it would be pointless if they can't get it out on time.

 

[Nereus77]

Intel is so far behind that they're about to be lapped by AMD.

This is like one of those "wasting time" fantasy questions that sports commentators come up with when one team is clearly smashing another team by a major margin on the field.

 

[mikk]

Intels Sunny Cove is 15% better clock per clock versus AMDs offerings and it supports AVX 512.

 

[Nereus77]

Intels Sunny Cove is 15% better clock per clock versus AMDs offerings and it supports AVX 512.

Sunny Cove will compete with the Ryzen 4000 series and I don't think Intel is ready for that....

 

[uzzi38]

Intels Sunny Cove is 15% better clock per clock versus AMDs offerings and it supports AVX 512.

Current offerengs yes (though it also clocks lower, so perf difference is less than 15%). On top of that, we're only, what, 1H away now until that flips on it's head? Whereas Intel's not really moving the needle by that point. Also lol at AVX512 support. Wake me up when it becomes relevant.

Slightly off topic, but you know what I feel worst for? ICL-SP. Completely dead in the water. Can't compete in core count, nor power efficiency, nor per-core performance, heck it's even later than Milan. No XCC die till Q1 2021 is a huge oof. Poor thing, died at the hand of Intel's ever constand delays and failed 10nm node. SPR is also just as dead in every regard.


Well, in any case, to answer OP: Ice Lake cores are significantly larger than Zen 2 cores and average about 15% better per clock performance, but they also guzzle more power and don't clock as high. ST performance I'd still put SNC ahead even given node failures, but mostly down to the fact that Zen 2 doesn't clock all that great. Assuming the vapourware 4.1GHz ICL-U here though, if we consider what actually exists on market, then Zen 2 is probably ahead. Got a frend who can sustain 4.6GHz ST on his 3950X in heavier workloads which do use some AVX, so just about better.

On the last point: none. They'd probably need to increase die sizes at some point even if they were to improve the Skylake uArch. They could try working on the node instead, squeeze out higher clocks, but given the use of TB3.0 and TVB on Comet Lake only pushing ST clocks up by 300mhz, I'd reckon they're at the limits of what can be extracted form 14nm too.

Unless they pull a Pentium IV. Drop some IPC for moar clocks.

 

[DrMrLordX]

Intels Sunny Cove is 15% better clock per clock versus AMDs offerings and it supports AVX 512.

That isn't even the fastest core that Intel has at their disposal. Willow Cove is ready for market and Golden Cove is in an unknown state of development (but probably pretty far along).

 

[uzzi38]

That isn't even the fastest core that Intel has at their disposal. Willow Cove is ready for market and Golden Cove is in an unknown state of development (but probably pretty far along).

Golden is a while away from hitting the market. Late 2021 is the best-case scenario, and that's assuming Alder Lake doesn't mysteriously drop off roadmaps (again) like Tiger Lake-S.

Willow Cove is a nice clock bump at least. Plus a really minor IPC bump, so there's that. It's also still nearly half a year away from hitting the market (give or take a month or two).

 

[turtile]

My assumption is that they will be on par after Zen 3 comes out. And at that point, AMD is in the clear win unless Intel changes the way that put together CPUs.

Intel has been designing their CPUs with their own process tech in mind. Now they are undoing that due to the difficulty with 10nm. Zen 2 had to move to TSMC last minute and was definitely not optimized for it.

 

[uzzi38]

Zen 2 had to move to TSMC last minute and was definitely not optimized for it.

Zen 2 was always TSMC, that's most certainly not the case. Zen 3 won't be a huge clock uplift, if that's what you're hoping for. Especially not on desktop.

 

[turtile]

Zen 2 was always TSMC, that's most certainly not the case. Zen 3 won't be a huge clock uplift, if that's what you're hoping for. Especially not on desktop.

Evidence?

The CPU design process is started four years before it hits physical form. Zen 1 was to be on GF 14nm which was canceled and replaced with Samsung 14nm. Zen 2 was to be on GF 7nm, then GF said they couldn't make the demand and made the design rules similar so they could share capacity with TSMC and then they just threw the entire 7nm out the window. AMD finally was forced to move Zen 2 to all TSMC 7nm.

Zen 3 will be mostly IPC uplift. Zen 4 and 5 will be the same. (actually, potential for clock loss)

 

[Ajay]

Evidence?

The CPU design process is started four years before it hits physical form. Zen 1 was to be on GF 14nm which was canceled and replaced with Samsung 14nm. Zen 2 was to be on GF 7nm, then GF said they couldn't make the demand and made the design rules similar so they could share capacity with TSMC and then they just threw the entire 7nm out the window. AMD finally was forced to move Zen 2 to all TSMC 7nm.

Zen 3 will be mostly IPC uplift. Zen 4 and 5 will be the same. (actually, potential for clock loss)

Nice story. Since you like evidence so much, please provide detailed citations so we can examine your claims.

 

[turtile]

Nice story. Since you like evidence so much, please provide detailed citations so we can examine your claims.

GF 14nm-XM that never came to be (2013):

GLOBALFOUNDRIES Accelerates Adoption of 20nm-LPM and 14nm-XM FinFET Processes with Comprehensive Production-Ready Design Flows

Jointly developed with leading EDA providers, flows address AMS challenges from specification to verification; complete flows for digital design for double patterning

www.globalfoundries.com

GF licenses Samsung (2014):

Samsung and GLOBALFOUNDRIES Forge Strategic Collaboration to Deliver Multi-Sourced Offering of 14nm FinFET Semiconductor Technology

Shared technology allows global capacity for 14nm FinFET fabrication in the U.S. and Korea

www.globalfoundries.com

GF 7nm that, again, never comes to be (June 2017):

“Our 7nm FinFET technology development is on track and we are seeing strong customer traction, with multiple product tapeouts planned in 2018,” said Gregg Bartlett, senior vice president of the CMOS Business Unit at GF.

GLOBALFOUNDRIES on Track to Deliver Leading-Performance 7nm FinFET Technology

New 7LP technology offers 40 percent performance boost over 14nm FinFET

www.globalfoundries.com

Right before news broke that there was no GF 7nm (July 2018):

So, Matt, on your first question relative to the manufacturing of the second generation of EPYC, so as I said earlier, we are working with both the TSMC and GLOBALFOUNDRIES in 7-nanometer. As for the 7-nanometer Rome that we're currently sampling, that's being manufactured at TSMC.

AMD “Rome” EPYC CPUs to Be Fabbed By TSMC

www.anandtech.com

So unless you think that AMD optimized for each process when they were design complete, I highly doubt that is the case...

 

[Ajay]

GF 14nm-XM that never came to be (2013):

GLOBALFOUNDRIES Accelerates Adoption of 20nm-LPM and 14nm-XM FinFET Processes with Comprehensive Production-Ready Design Flows

Jointly developed with leading EDA providers, flows address AMS challenges from specification to verification; complete flows for digital design for double patterning

www.globalfoundries.com

GF licenses Samsung (2014):

Samsung and GLOBALFOUNDRIES Forge Strategic Collaboration to Deliver Multi-Sourced Offering of 14nm FinFET Semiconductor Technology

Shared technology allows global capacity for 14nm FinFET fabrication in the U.S. and Korea

www.globalfoundries.com

GF 7nm that, again, never comes to be (June 2017):

GLOBALFOUNDRIES on Track to Deliver Leading-Performance 7nm FinFET Technology

New 7LP technology offers 40 percent performance boost over 14nm FinFET

www.globalfoundries.com

Right before news broke that there was no GF 7nm (July 2018):

AMD “Rome” EPYC CPUs to Be Fabbed By TSMC

www.anandtech.com

So unless you think that AMD optimized for each process when they were design complete, I highly doubt that is the case...

Sorry, I meant this "Zen 3 will be mostly IPC uplift. Zen 4 and 5 will be the same. (actually, potential for clock loss) ". The others I was aware of.
Thanks for putting in the effort though!

 

[HurleyBird]

If we could get both companies to design a CPU core now on the same process and using the same number of transistors, who would be ahead and by how much? I'm assuming Intel would be ahead because Skylake has a similar IPC to Zen 2, but Zen 2 uses a more dense node, and Skylake is from 2015 while Zen 2 is from 2019.

The CPU core is only half of the equation. The uncore is also pretty important, where Intel's uncore seems better for pure performance (significantly better latency), while AMD's has a massive advantage in manufacturability and scalability (hence why Rome is so darn good). Take the Zen 2 core and couple it with an Intel-type uncore, and Sunny Cove's IPC advantage might shrink by a good amount. Certainly gaming would see a large boost.

 

[Ajay]

The CPU core is only half of the equation. The uncore is also pretty important, where Intel's uncore seems better for pure performance (significantly better latency), while AMD's has a massive advantage in manufacturability and scalability (hence why Rome is so darn good). Take the Zen 2 core and couple it with an Intel-type uncore, and Sunny Cove's IPC advantage might shrink by a good amount. Certainly gaming would see a large boost.

That's true now, but Intel will need to move towards some sort of chiplet based CPU system in order to expand core counts and maintain economical production on the ever expanding cost of manufacturing wafers on smaller nodes.

 

[turtile]

Sorry, I meant this "Zen 3 will be mostly IPC uplift. Zen 4 and 5 will be the same. (actually, potential for clock loss) ". The others I was aware of.
Thanks for putting in the effort though!

Well, there is a whole presentation by Forrest Norrod:

 

[DrMrLordX]

Golden is a while away from hitting the market. Late 2021 is the best-case scenario, and that's assuming Alder Lake doesn't mysteriously drop off roadmaps (again) like Tiger Lake-S.

Willow Cove is a nice clock bump at least. Plus a really minor IPC bump, so there's that. It's also still nearly half a year away from hitting the market (give or take a month or two).

Willow Cove is ready. There are samples that have turned up in Geekbench and elsewhere. They're just bugfixing the silicon at this point. In light of what the OP asked, should Intel just magically *poof* a CPU into existence today on whatever process they wanted (which admittedly, requires respins to follow the design rules of said process), Willow Cove would be ready to go, no doubt.

Golden Cove, we don't know. Intel had Cannonlake ready in 2017 (and possibly earlier; we don't know how many of the Cannonlake delays were due to problems with early 10nm silicon). Intel followed up with IceLake in 2019, but IceLake/Sunny Cove was probably ready in simulations as far back as 2018, unless you believe it took Intel more than a year to do the design followup in their labs (I don't). Presumably Willow was ready last year in simulation, and Golden probably exists now (or will be finalized sometime this year).

Make no mistake: Intel hasn't stood still in their labs. The designs are there (rumours about Ocean Cove notwithstanding). They just can't produce ES chips with acceptable clocks and defect rates, which is why it's taking them so long to produce marketable product with said designs. In the real world, targeting a different process is pretty time-intensive which means it's not like Intel can take Willow or Golden and produce it at TSMC or Samsung today. So as long as Intel has problems with their processes and as long as they continue to tow the company line in targeting only their own processes, we'll continue to see massive lag between R&D and actual product. Even if they change direction now, it'd be a year at least before they could take existing designs (presumably Willow) and port it to a Samsung process.

 

[yuri69]

Intel's recent designs has higher IPC and lower latency than AMD's Zen2. However, they are unable to manufacture them for desktop and server.

It's interesting that even if they had manufactured Ice Lake-SP on plan early 2019 or so. It would most probably feature just 38c. So even with its ~15% IPC advantage, it would still have to face the 64c Rome.

 

[uzzi38]

Willow Cove is ready. There are samples that have turned up in Geekbench and elsewhere. They're just bugfixing the silicon at this point. In light of what the OP asked, should Intel just magically *poof* a CPU into existence today on whatever process they wanted (which admittedly, requires respins to follow the design rules of said process), Willow Cove would be ready to go, no doubt.

Golden Cove, we don't know. Intel had Cannonlake ready in 2017 (and possibly earlier; we don't know how many of the Cannonlake delays were due to problems with early 10nm silicon). Intel followed up with IceLake in 2019, but IceLake/Sunny Cove was probably ready in simulations as far back as 2018, unless you believe it took Intel more than a year to do the design followup in their labs (I don't). Presumably Willow was ready last year in simulation, and Golden probably exists now (or will be finalized sometime this year).

Make no mistake: Intel hasn't stood still in their labs. The designs are there (rumours about Ocean Cove notwithstanding). They just can't produce ES chips with acceptable clocks and defect rates, which is why it's taking them so long to produce marketable product with said designs. In the real world, targeting a different process is pretty time-intensive which means it's not like Intel can take Willow or Golden and produce it at TSMC or Samsung today. So as long as Intel has problems with their processes and as long as they continue to tow the company line in targeting only their own processes, we'll continue to see massive lag between R&D and actual product. Even if they change direction now, it'd be a year at least before they could take existing designs (presumably Willow) and port it to a Samsung process.

1. By that definition Zen 3 is also ready, and has been for months.

2. You're right, we don't know about Golden. But it'll be the same as every core every core - it's finalized long over a year prior to release. Zen 4 will be finalised mid this year by that measure (providing enough time for initial tapeouts, sampling to select customers prior to the end of the year). Back to Golden - you can be certain they're tweaking and tuning it until the last moment, even if it's not the main priority.

3. Agreed on the last point for the most part.

 

[Adonisds]

In light of what the OP asked, should Intel just magically *poof* a CPU into existence today on whatever process they wanted (which admittedly, requires respins to follow the design rules of said process), Willow Cove would be ready to go, no doubt.

I didn't ask that at all. It seems some people here couldn't abstract the question.

It's not a question asking who is ahead now and what could Intel do. Few could reasonably disagree that AMD right now has better products.

It's also not asking what architecture right now has better IPC. The data is pretty clear that Sunny Cove has better IPC than Zen 2, but Ice Lake has yield and clock problems, etc...

I just wanted to know how important is CPU design. Do two different teams, one with more resources than the other, produce nearly the same CPUs if they target the same ISA and process? Is what's going to define what CPU core will perform better mostly what company is using the more advanced process?

Sunny Cove has around 15% better IPC than Zen 2. But that doesn't mean the design is necessarily 15% better. If it's using 80% more transistors to deliver 15% more performance it's a worse design. That's the kind of comparison I hoped someone would do.

 

[TheGiant]

Willow Cove is ready. There are samples that have turned up in Geekbench and elsewhere. They're just bugfixing the silicon at this point. In light of what the OP asked, should Intel just magically *poof* a CPU into existence today on whatever process they wanted (which admittedly, requires respins to follow the design rules of said process), Willow Cove would be ready to go, no doubt.

Golden Cove, we don't know. Intel had Cannonlake ready in 2017 (and possibly earlier; we don't know how many of the Cannonlake delays were due to problems with early 10nm silicon). Intel followed up with IceLake in 2019, but IceLake/Sunny Cove was probably ready in simulations as far back as 2018, unless you believe it took Intel more than a year to do the design followup in their labs (I don't). Presumably Willow was ready last year in simulation, and Golden probably exists now (or will be finalized sometime this year).

Make no mistake: Intel hasn't stood still in their labs. The designs are there (rumours about Ocean Cove notwithstanding). They just can't produce ES chips with acceptable clocks and defect rates, which is why it's taking them so long to produce marketable product with said designs. In the real world, targeting a different process is pretty time-intensive which means it's not like Intel can take Willow or Golden and produce it at TSMC or Samsung today. So as long as Intel has problems with their processes and as long as they continue to tow the company line in targeting only their own processes, we'll continue to see massive lag between R&D and actual product. Even if they change direction now, it'd be a year at least before they could take existing designs (presumably Willow) and port it to a Samsung process.

after a long time, we can not find a icelake server CPU leaks
it should be out IMO, but it isn't...