64 core EPYC, first benchmark leaked

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,483
14,434
136

EXCellR8

Diamond Member
Sep 1, 2010
3,979
839
136
Well it's 7nm... idk could be credible to some extent. No info on the board used?
 

Abwx

Lifer
Apr 2, 2011
10,854
3,298
136
This amount to 196pts/Core, about what a R7 2700 does currently, so the score is quite possible given that Zen 2 is supposed to bring slightly higher throughput/Hz.
 

inf64

Diamond Member
Mar 11, 2011
3,685
3,957
136
the score is quite possible given that Zen 2 is supposed to bring slightly higher throughput/Hz.
Slightly higher khm khm :D

PS Not saying I believe that this score is real, it looks like a fake.
 

BigDaveX

Senior member
Jun 12, 2014
440
216
116
This amount to 196pts/Core, about what a R7 2700 does currently, so the score is quite possible given that Zen 2 is supposed to bring slightly higher throughput/Hz.
I dunno, it seems kinda unlikely that what is presumably a fairly early engineering sample would clock on-par with a chip with eight times fewer cores - for perspective the 2700 has 3.2GHz base / 4.1GHz turbo clocks, while the fastest 32C Epyc is clocked at 2.2GHz / 2.7GHz - unless the 7nm process the chip is using completely blows the doors off any other process currently out there.
 

dnavas

Senior member
Feb 25, 2017
355
190
116
I dunno, it seems kinda unlikely that what is presumably a fairly early engineering sample would clock on-par with a chip with eight times fewer cores - for perspective the 2700 has 3.2GHz base / 4.1GHz turbo clocks, while the fastest 32C Epyc is clocked at 2.2GHz / 2.7GHz - unless the 7nm process the chip is using completely blows the doors off any other process currently out there.

It's aggressive for sure. Let's see, Huawei quoted its mobile 7nm processor as being 40% more efficient AND 20% faster. If Rome scales the same, speed would be 2.6G and power would be 180W * 2 * .6. Call that 220W. Perf is still about 60% away from target. Maybe SMT4 brings 30% uplift in CB? That would mean an effective SMT2 speed of 3.4G. You'd still need about a 20% uplift in IPC or freq, if I've done my math right. AMD would have to hit on literally every cylinder. In an ES?!

That said, Huawei's comparison was to 10nm (Kirin), which is supposed to be 15% higher than TSMC 16nm and 35% more efficient. So maybe? If this is even remotely true, this is about to be astonishing. Almost the way new silicon worked two decades ago.

Yeah, no, those days are over. I'm going to go with Stilt on this one -- a 2P system running at just under 2G. That would make more sense. It's the sort of system that's likely to be benched -- why bench 1P?
 

BigDaveX

Senior member
Jun 12, 2014
440
216
116
That said, Huawei's comparison was to 10nm (Kirin), which is supposed to be 15% higher than TSMC 16nm and 35% more efficient. So maybe? If this is even remotely true, this is about to be astonishing. Almost the way new silicon worked two decades ago.
Yeah, the last time we saw a new process that brought a major clockspeed improvement to existing CPU designs was probably AMD's 45nm jump, which was just shy of a decade ago. Even with them moving away from GloFo, I have a hard time believing any process could produce that big of a leap.
 
  • Like
Reactions: pcp7

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
If it really is Rome, then we're looking at Zen 2, not Zen/Zen+, so it's impossible for us to do direct clockspeed comparisons. Also if it is an engineering sample, it may not be running at production clocks. If Intel can run their chips with chillers then so can AMD.
 
  • Like
Reactions: ryan20fun

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,483
14,434
136
If it really is Rome, then we're looking at Zen 2, not Zen/Zen+, so it's impossible for us to do direct clockspeed comparisons. Also if it is an engineering sample, it may not be running at production clocks. If Intel can run their chips with chillers then so can AMD.
I am running my 2990wx at barely over stock vcore, and 3700. So Rome could very well do 3.7-3.8 stock, and if the IPC were even 7% better, we are right in line with this being real. I know that's a lot of assumptions, but just saying, based on mine, I say its possible.
 

Abwx

Lifer
Apr 2, 2011
10,854
3,298
136
It's aggressive for sure. Let's see, Huawei quoted its mobile 7nm processor as being 40% more efficient AND 20% faster.

40% more efficient AND 20% faster means 2x more efficient at same frequency, and that s without accounitng for throughput/Hz improvement wich allow to dial down frequency and increase perf/watt in the same ratio, so how it is presented your maths are somewhat off by quite a margin..

FTR a 64C using Pinnacle Ridge would be below 400W/3GHz..
 

maddie

Diamond Member
Jul 18, 2010
4,723
4,628
136
40% more efficient AND 20% faster means 2x more efficient at same frequency, and that s without accounitng for throughput/Hz improvement wich allow to dial down frequency and increase perf/watt in the same ratio, so how it is presented your maths are somewhat off by quite a margin..

FTR a 64C using Pinnacle Ridge would be below 400W/3GHz..
Yes, with that power saving, you can run double the cores at the same speed and less power.
 

StinkyPinky

Diamond Member
Jul 6, 2002
6,761
777
126
No one has any idea of the IPC gains in Zen 2 so not sure how anyone can possible comment to say it is fake or real. We are so used to Intels paltry gains in the last decade, but perhaps AMD have found a bit more juice for their next gen cpus.
 
  • Like
Reactions: ryan20fun

Asterox

Golden Member
May 15, 2012
1,026
1,775
136
It is either totally fake or alternatively a 2P system.
A 64-core CPU won't be hitting frequencies high enough to produce that score.

In case it is real, it is probably 2P with each (of the 128) core(s) running at < 2GHz.

Well we now that this CPU exist, or it is first AMD CPU product on TSMC 7nm.

Ok this is Trailer for "old Epyc CPU-s", "but 64 Cores/ 128 Threads this is to much performanse and options etc".:D

 

The Stilt

Golden Member
Dec 5, 2015
1,709
3,057
106
Well we now that this CPU exist, or it is first AMD CPU product on TSMC 7nm.

Ok this is Trailer for "old Epyc CPU-s", "but 64 Cores/ 128 Threads this is to much performanse and options etc".:D


I have no doubts that a 64-core EPYC will exist sooner or later, but rather if the alleged leak is even remotely accurate in terms of the used hardware (or at all).
The existing SP3 boards (which must be compatible with the upcoming Rome SKUs as well), have been designed for 200W max. sustained power draw (180W TDP with cTDP upconfig).
Even if the alleged 64C SKU had 50% higher power rating (i.e. 300W) and provided had >= 10% higher IPC in Cinebench R15 specifically, it still would need to have > 83% higher power efficiency than the parts made on 12nm LP.
Not very likely, as the CPU should be running closer to 4GHz to produce that kind of a score.

If there is any truth in the alleged leak, it is most certainly a 2P (i.e. 2x 64C/128T) system running at low speeds.
Being real would also mean that it is a rather early prototype part, which tend to run at even low(er) speeds than the generally low clocked high core count server parts already do.

In case you think it is real deal (which I personally don't), ask yourself which is more likely: An early prototype 64C/128T CPU running at < 4GHz or two early prototype 64C/128T CPUs running at < 2GHz in a 2P system?

While I have no doubts regarding the superiority of the 7nm process from TSMC compared to GF 12/14nm processes, I cannot see the supplied information adding up even remotely considering the realistic characteristics of a server CPU and the existing infrastructure especially.

It is also becoming a habit that fake AMD news gets "leaked" when Intel is close to releasing something new. When 8700K was to be released we were seeing "news" about Ryzen 5 series CPUs shipping with all 8 cores enabled. Now that Coffee Lake Refresh is about to be released, we've already seen "Ryzen 7 2800X with 10 cores" and now this.
 
Last edited:

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
It's aggressive for sure. Let's see, Huawei quoted its mobile 7nm processor as being 40% more efficient AND 20% faster.

In the semiconductor industry, "AND" means or. So its 40% more efficient or 20% faster. Anandtech wrote a detailed article about the chip: https://www.anandtech.com/show/13298/hisilicon-announces-the-kirin-980-first-a76-g76-on-7nm

You can also see from the ARM graph how such claims work. 40% faster in same power envelope, 50% lower power at same performance.

Top-to-top comparisons are also shown in HiSilicon's own chart. It claims 37% faster in Geekbench 4, but 32% more efficient in Dhrystone. As flawed as Geekbench may be, its not as bad as Dhrystone.

So am I saying there's no gain for 7nm EPYC? Of course not! Going from 32 cores to 64 cores at the same TDP is a huge gain. 2P also happens to be a standard testing configuration for server chips so Stilt's assumptions are likely accurate too.
 
  • Like
Reactions: dnavas

LightningZ71

Golden Member
Mar 10, 2017
1,627
1,898
136
So, are we maybe looking at 64 cores with SMT disabled? We've seen that SMT can lead to greater power usage in some chips, and negatively influence over clocking stability. Perhaps this is with a single 64 cores package with SMT off, pushing maximum frequencies within the power envelope available.

This may be an AMD key competitive advantage. With SMT implementations increasingly under attack via timing and side channel attacks, being able to provide maximum core numbers at high sustained clocks could be something that AMD can leverage for market advantage. It would certainly help manage the massive memory bandwidth demands that 128 threads would have.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
So, are we maybe looking at 64 cores with SMT disabled? We've seen that SMT can lead to greater power usage in some chips,

It would certainly help manage the massive memory bandwidth demands that 128 threads would have.

SMT is 30% faster for less than 30% power increase. That's a far more power efficient feature than most architectural changes. It also increases power used because it increases utilization, which is what you want.

Also part of the reason SMT provides the increase is because it makes it reduces the impact of DRAM being so far away from the CPU by hiding latency.

They aren't going to disable such a performance enhancing feature. They'll find a way to fix bugs and continue to use them.
 

iamgenius

Senior member
Jun 6, 2008
803
80
91
Here: https://www.techradar.com/news/leak-shows-amds-64-core-epyc-processor-blowing-away-intels-xeon-cpus

One small excerp:
"The benchmark was highlighted by Chiphell (as spotted by Techquila) and shows how the 64-core CPU performs in Cinebench R15’s multi-threaded test, achieving a colossal score of 12,587."

That more than double of my 32 core 2990wx overclocked to 4.0 getting about 6000
That's epyc(epic). Assuming it is real, how much do you think AMD will charge for it?
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,483
14,434
136
So, are we maybe looking at 64 cores with SMT disabled? We've seen that SMT can lead to greater power usage in some chips, and negatively influence over clocking stability. Perhaps this is with a single 64 cores package with SMT off, pushing maximum frequencies within the power envelope available.

This may be an AMD key competitive advantage. With SMT implementations increasingly under attack via timing and side channel attacks, being able to provide maximum core numbers at high sustained clocks could be something that AMD can leverage for market advantage. It would certainly help manage the massive memory bandwidth demands that 128 threads would have.
From what I read, its HT that has the problem, not AMD's SMT.
 
  • Like
Reactions: Drazick

Asterox

Golden Member
May 15, 2012
1,026
1,775
136
I have no doubts that a 64-core EPYC will exist sooner or later, but rather if the alleged leak is even remotely accurate in terms of the used hardware (or at all).
The existing SP3 boards (which must be compatible with the upcoming Rome SKUs as well), have been designed for 200W max. sustained power draw (180W TDP with cTDP upconfig).
Even if the alleged 64C SKU had 50% higher power rating (i.e. 300W) and provided had >= 10% higher IPC in Cinebench R15 specifically, it still would need to have > 83% higher power efficiency than the parts made on 12nm LP.
Not very likely, as the CPU should be running closer to 4GHz to produce that kind of a score.

If there is any truth in the alleged leak, it is most certainly a 2P (i.e. 2x 64C/128T) system running at low speeds.
Being real would also mean that it is a rather early prototype part, which tend to run at even low(er) speeds than the generally low clocked high core count server parts already do.

In case you think it is real deal (which I personally don't), ask yourself which is more likely: An early prototype 64C/128T CPU running at < 4GHz or two early prototype 64C/128T CPUs running at < 2GHz in a 2P system?

While I have no doubts regarding the superiority of the 7nm process from TSMC compared to GF 12/14nm processes, I cannot see the supplied information adding up even remotely considering the realistic characteristics of a server CPU and the existing infrastructure especially.

It is also becoming a habit that fake AMD news gets "leaked" when Intel is close to releasing something new. When 8700K was to be released we were seeing "news" about Ryzen 5 series CPUs shipping with all 8 cores enabled. Now that Coffee Lake Refresh is about to be released, we've already seen "Ryzen 7 2800X with 10 cores" and now this.

My comment goes only for 64 Cores/128 Threds Epyc CPU is live and kicking.Everything else is not important, it is server CPU or "red server CPU Rambo is only make sense in Twilight Zone".