Speculation: Ryzen 4000 series/Zen 3

Page 76 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

amd6502

Senior member
Apr 21, 2017
971
360
136
How exactly do you know this? Does throughput increase less than power with SMT?

Perf/Watt should decrease without SMT.

However, sparse thread and single thread performance should see advantages, as boost can be sustained longer and higher frequencies can be hit.

With 7nm being so temperature sensitive, and power density and local heat also an issue, a few idle execution units on the average are a very good thing if all you care about is high frequencies.

Maybe most importantly , there is never unintentional loss of performance to thread from the OS tasksetting it to an occupied core (In real life, misconfigurations and suboptimal execution is probably the norm rather than the exception), as well as a higher consistency of what a thread gets in terms of resources, including caches such as the L1 and L2. .
 
  • Like
Reactions: Space Tyrant

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,575
146
Perf/Watt should decrease without SMT.

However, sparse thread and single thread performance should see advantages, as boost can be sustained longer and higher frequencies can be hit.

With 7nm being so temperature sensitive, and power density and local heat also an issue, a few idle execution units on the average are a very good thing if all you care about is high frequencies.

Maybe most importantly , there is never unintentional loss of performance to thread from the OS tasksetting it to an occupied core (In real life, misconfigurations and suboptimal execution is probably the norm rather than the exception), as well as a higher consistency of what a thread gets in terms of resources, including caches such as the L1 and L2. .
You're right, it should.

Should.

So we tested this shortly after the 4700U 8c8t leaks a while back, and a few of us on a Discord chat nearly went insane with the results.

Here's a write up:
So there's a possibility that SMT disabled improves perf/watt. As insane as that sounds.
 

Veradun

Senior member
Jul 29, 2016
564
780
136
Yeah, should or shouldn't my point is an 8c/8t setup is perfectly in line with what Devs have been using the past generation and doesn't bear any additional baggage for additional optimization steps.
 

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
Yeah, should or shouldn't my point is an 8c/8t setup is perfectly in line with what Devs have been using the past generation and doesn't bear any additional baggage for additional optimization steps.
I'd still be shocked if they shipped them without SMT. I can't know for sure of course.
 

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,575
146
I'd still be shocked if they shipped them without SMT. I can't know for sure of course.

The final SKU list as announced by AMD does come with SMT disabled skus. The 4300U, 4500U and 4700U all are SMT-less.
 

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
The final SKU list as announced by AMD does come with SMT disabled skus. The 4300U, 4500U and 4700U all are SMT-less.
I thought we were talking about the new consoles. What I wrote, I meant for the consoles.
 

Richie Rich

Senior member
Jul 28, 2019
470
229
76
This is very interesting part from AMD's CEO interview:

AnandTech: .... Arm has put forward a roadmap for its leading enterprise cores to deliver performance gains of 25% year on year, which goes beyond AMD’s ‘above market’ target of 7%. Is there a time where AMD would look into doing a high-performance Arm design?
LS: .... From a server standpoint, we are not investing in Arm at this point. We think that there’s a huge market out there for x86. I do think Arm has a market and capability, but from our standpoint the focus on x86 is the right thing to do. ....

VentureBeat: Looking at the whole industry, what do you think of RISC-V as a challenger to the way of doing things?
LS: There will be those that use it, and they've seen some good momentum. Our focus is very clear, it’s all about high-performance compute, and for that x86 is the leader.

Next time Anandtech should ask why AMD's fastest desktop CPU the mighty 16-core RYZEN 9 3950X clocked at 4.7 Ghz (highly selected silicon) is slower in single thread performance than every iPhone 11. In other words why Apple core is 82% faster at same clock speed than AMD high performance chip.

I'm pretty sure that AMD CEO cannot officially say something like this: "Yeah we did big mistake to shut down K12 project and let Mr. Keller leave." At least I hope AMD management know the real IPC numbers and plan future development accordingly. Because Zen2 brought 15% IPC after two years and that's just 7.5% IPC per year. Cortex A77 matched (in SPEC exceeded Zen2) performance however ARMs 25% plan is 3x faster development than AMD's. Hopefully they cancelled K12 ARM branch only (6xALU core+SMT) and its x86 branch will be Zen 3 (both based on same back-end core so Zen 3 is 6xALU core too).

 

HutchinsonJC

Senior member
Apr 15, 2007
465
201
126
Isn't it like this?

Going Wider = harder to get higher frequency

Higher Frequency = harder to go wider

six of one... half a dozen of the other?
 

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
This is very interesting part from AMD's CEO interview:



Next time Anandtech should ask why AMD's fastest desktop CPU the mighty 16-core RYZEN 9 3950X clocked at 4.7 Ghz (highly selected silicon) is slower in single thread performance than every iPhone 11. In other words why Apple core is 82% faster at same clock speed than AMD high performance chip.

I'm pretty sure that AMD CEO cannot officially say something like this: "Yeah we did big mistake to shut down K12 project and let Mr. Keller leave." At least I hope AMD management know the real IPC numbers and plan future development accordingly. Because Zen2 brought 15% IPC after two years and that's just 7.5% IPC per year. Cortex A77 matched (in SPEC exceeded Zen2) performance however ARMs 25% plan is 3x faster development than AMD's. Hopefully they cancelled K12 ARM branch only (6xALU core+SMT) and its x86 branch will be Zen 3 (both based on same back-end core so Zen 3 is 6xALU core too).

Ah here we go again... can somebody else please do the usual rounds? I got so much brain fatigue from this guy.
 

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
AMD doesn't really have to care about Apple's SoC performance. AMD doesn't sell CPUs to Apple as is so they can't be replaced with Apple's chips there. Apple doesn't sell to other third parties that AMD does, so they don't compete there either.

No one else in the ARM space is even close to Apple in terms of performance. Also there's no real information about how high Apple can clock their design. Some of the choices they've made no doubt mean that you can't actually get their SoC to 4.7 GHz or perhaps even close to it.
 
Feb 4, 2009
34,506
15,737
136
And she could certainly answer:

"Ian, are you running AnandTech website on Xeon or iPhone 11s, and why?"

I know, one is made to run a very specialized os and app, the other is made to run on multiple os and multiple apps.
I would bet there is an escalator control chip out there somewhere that has huge single thread performance
 
  • Haha
  • Like
Reactions: Tlh97 and lobz

Richie Rich

Senior member
Jul 28, 2019
470
229
76
AMD doesn't really have to care about Apple's SoC performance. AMD doesn't sell CPUs to Apple as is so they can't be replaced with Apple's chips there. Apple doesn't sell to other third parties that AMD does, so they don't compete there either.

No one else in the ARM space is even close to Apple in terms of performance. Also there's no real information about how high Apple can clock their design. Some of the choices they've made no doubt mean that you can't actually get their SoC to 4.7 GHz or perhaps even close to it.
I agree with you those are separated markets because Apple will never enter server market. However just imagine you would be CEO of AMD and you fight with Intel for every single digit of IPC %. This means when you are slower than Intel more than 10% you will have problem with sales, slower more than 20% means loosing market shares and margins, slower more than 40% (Bulldozer era) means you are going to bankruptcy. Wouldn't you as CEO be a little nervous when somebody has CPU with +80% IPC advantage? More than double what you consider as bankruptcy level. Honestly I would be VERY nervous. At least due to possibility when chief architect would leave Apple to your competitor (and give them instant 4 years of CPU development advantage). Most of the architectural tricks in Apple core could be applied to x86 as well (because x86 CISC is internaly RISC running similar back end as ARM).

Lisa Su could be calm only under condition of Zen 3 is Keller's K12 (x86 branch) with 6xALUs. If this is true a lot of people will be surprised by Zen 3 performance. Something like those leaks about 40-50% IPC uplift is true, but it's uplift everywhere (int + FPU) and not just FPU.
 
  • Haha
Reactions: CHADBOGA

moinmoin

Diamond Member
Jun 1, 2017
4,934
7,619
136
AMD is focusing on x86. The server market is dominated by x86. Zen is x86. In these circumstances the only competitor AMD has to worry about is Intel (and potentially VIA/Centaur).
 

Richie Rich

Senior member
Jul 28, 2019
470
229
76
Is Frequency NOT part of the performance of the chip?
Yes, frequency is part of the performance of course. The problem is designing 4 GHz CPU is much much simpler than designing CPU with double IPC. Scalar IPC is the holy grale of CPU architecture, the most dificult thing to develop. Big problem is not Apple but company like Nuvia - ex-Apple CPU architects are developing new core with Apple IPC and high clocks around 4Ghz. In 2024 Nuvia will start shipping their new CPU for servers and that's a problem for AMD. They need to uplift IPC by 80% in just 4 years otherwise they will be in big trouble.

AMD is focusing on x86. The server market is dominated by x86. Zen is x86. In these circumstances the only competitor AMD has to worry about is Intel (and potentially VIA/Centaur).
That's true nowadays. But x86 dominates server market due to high-performance with lower price it brought in. Don't forget there was many dominant server ISAs in history (DEC Alpha, Sparc, PowerPC). And x86 killed them all by performance. Do you think x86 will dominate market even though ARM will provide cheaper and more powerfull CPUs?
 
Last edited:

gdansk

Golden Member
Feb 8, 2011
1,979
2,355
136
Wouldn't you as CEO be a little nervous when somebody has CPU with +80% IPC advantage?
In micro-benches where all you're testing is if the workload fits in their massive L1 and L2 caches. Most real work is memory bandwidth constrained. And AMD and Intel have between a 70 and 100% frequency advantage. You think frequency scaling is something easy, it isn't. AMD and Intel have spent a great deal of time and money reworking their designs to clock high. Scaling beyond a few high performance cores is another issue, the interconnects sap power and it's part of the reason server-focused designs use L3 cache.

Another reason AMD and Intel are not afraid on servers is because they have run simulations of their chips with much larger L1 and L2 caches and determined that their IPC rockets up when SPECint2006 benches fit in the cache. It's hard to justify the die space, though, when you're trying to build a 64 core design.
 
Last edited:

moinmoin

Diamond Member
Jun 1, 2017
4,934
7,619
136
Do you think x86 will dominate market even though ARM will provide cheaper and more powerfull CPUs?
For quite some time, yes. The slow rise of AMD's server market share shows the uphill battle ARM faces double fold.

And I believe AMD won't be unprepared if ARM makes bigger inroads. ARM is already a big part of AMD's tech. But as one of two duopolistic manufactures of x86 chips AMD has no interest whatsoever to deteriorate their own market.
 

scannall

Golden Member
Jan 1, 2012
1,944
1,638
136
This is very interesting part from AMD's CEO interview:

Next time Anandtech should ask why AMD's fastest desktop CPU the mighty 16-core RYZEN 9 3950X clocked at 4.7 Ghz (highly selected silicon) is slower in single thread performance than every iPhone 11. In other words why Apple core is 82% faster at same clock speed than AMD high performance chip.

The answer has nothing to do with hardware. And everything to do with software installed base. If you came up with a magic CPU that was %500 percent faster clock for clock it would still take 15 or 20 years to make x86 irrelevant. Maybe longer. So, for now and the foreseeable future x86 is where the money is.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
In other words why Apple core is 82% faster at same clock speed than AMD high performance chip.

You really going to Geekbench* the number eh?

In the real world it doesn't scale 100%, otherwise there wouldn't be any focus on improving the memory subsystem or the I/O.

Some like HPC, scale even less than that. SpecFP is representative, and the scaling factor is often 60-70%.

(Geekbench scales pretty much 100% with clocks, so do many less realistic benchmarks and also synthetics)