Alleged AMD EPYC 7003 128-Core 2P Zen 3 System Cranks Insane Cinebench R23 Score

erek

erek

[H]F Junkie
Joined
Dec 19, 2005
Messages
10,891
"All that parallelism should give AMD a leg up on very wide CPU loads that require a lot of hardware threads. Zen 3's improved single-thread performance will no doubt bolster EPYC's performance as well. Last month, Intel compared its new Ice Lake-based 32-core Xeon processors to the 64-core EPYC 7742 and claimed that the Xeons thumped AMD despite having fewer cores. Most of those loads required the higher memory bandwidth afforded by Xeons and the AVX-512 instruction set. That situation hasn't changed with Zen 3, so Intel might still have a few edge case wins. Regardless, AMD looks like it will be even more competitive in the server space once Milan EPYC CPUs launch."

https://hothardware.com/news/128-core-zen-3-epyc-cinebench
 
Competitive only if the big players sell servers with the chips in them....Dell/HPE/Cisco et cetera
 
Mega6 said:
The "old guard" dumb ass corporate purchasers will just continue to buy intel because "nobody ever got fired for buying intel". Hopefully they get fired.
Click to expand...
I can't emphasis this exact point enough - agreed.
AMD has picked up the slack in literally every single market and area that Intel has completely failed in.

Intel has still yet to make up for the severe performance decreases and never ending security risks that have emerged over the last two years.
From the article:
According to ExecutableFix, such a system can achieve a Cinebench R23 score of 87,878. Unfortunately most of our Cinebench scores in recent reviews are for version R20, but we do have one point of comparison. Apple's octo-core M1 Mac mini topped out at 7665, so the EPYC server was around 11.5x faster than Apple's tiny desktop CPU. In other words, the EPYC 7763 appears to have parallelism in spades.
Click to expand...
AMD is the only megacorp keeping x86-64 alive, and with the continued results I am seeing, and if they can keep up this momentum, then I retract my statements about x86-64 being all but dead within a decade.
ARM development is going to have an uphill battle with AMD's x86-64, and the silver-lining is, competition is good. (y)
 
Lakados said:
The Two I bought from Dell in July didn’t arrive until mid Oct?
Click to expand...

Micro experience.

Just configured a top of the line R7525 with dual Milan processors and it gave me an estimated shipping date of 11JAN
Screenshot_20201213-203115.png
 
alxlwson I live in a world where companies still only buy intel based servers and cisco networking gear because they always have and always will. That is the majority of Enterprise Medium to large businesses still to this day. Trying to convince them to go with AMD or anything else is like pulling teeth.
 
According to ExecutableFix, such a system can achieve a Cinebench R23 score of 87,878. Unfortunately most of our Cinebench scores in recent reviews are for version R20, but we do have one point of comparison. Apple's octo-core M1 Mac mini topped out at 7665, so the EPYC server was around 11.5x faster than Apple's tiny desktop CPU. In other words, the EPYC 7763 appears to have parallelism in spades.
Click to expand...

11.5 times faster with 16 times the cores? Apple is gonna smoke Zen3 if they can scale up. Intel's not even in the room any more.
 
MrGuvernment said:
alxlwson I live in a world where companies still only buy intel based servers and cisco networking gear because they always have and always will. That is the majority of Enterprise Medium to large businesses still to this day. Trying to convince them to go with AMD or anything else is like pulling teeth.
Click to expand...
There is something to be said for sticking with what works. Don't get me wrong I love my AMD Epycs, even the little Embedded one I have as a test bench but they haven't been without their growing pains. And AMD's Bios updates break as many things as they fix but once you get past those issues they make great virtualization servers. But I still think that Intel can deliver the better overall Package, AMD has the best CPU for sure, but drivers, storage, networking, and many of the finer points they are still behind, but I will welcome anything that can bring Intel in line or that give alternatives.
 
Last edited:
idiomatic said:
11.5 times faster with 16 times the cores? Apple is gonna smoke Zen3 if they can scale up. Intel's not even in the room any more.
Click to expand...

Not only that, a full-fledged server/workstation product where TDP and noise doesn't really matter is only 11.5 times faster with 16 times the cores than a fanless, mobile, generation 1 processor that's in a $999 ultrabook.

I am both an AMD and an Apple fan - I have a 3990x number cruncher and a loaded MBP, both of which I am insanely happy with. I am REALLY looking forward to the exponential CPU gains we are about to see over the next five years.
 
idiomatic said:
11.5 times faster with 16 times the cores? Apple is gonna smoke Zen3 if they can scale up. Intel's not even in the room any more.
Click to expand...
It is a big if thought right.

For example we could have used the exact same logic with zen 3, a 5800x score 15,245 on R23 and a 8 core 4800U is also doing much better than 88K / 16

Epyc is just 5.76 times faster than the 5800x while having 16 times more core, AMD is gonna smoke the Zen3 if the Zen3 can scale up.

The M1 does 1498 points in single core with the macbook pro and 7508 with all of them according to this:
https://www.notebookcheck.net/These...impressive-for-a-10-W-processor.504380.0.html

Scaling significantly less than Zen3 (I imagine that is because they have 4 performance core + 4 non full core ?) So maybe in fact they scale better and it would show if they would have 8 full performance core.

paradoxical said:
Not only that, a full-fledged server/workstation product where TDP and noise doesn't really matter is only 11.5 times faster with 16 times the cores than a fanless, mobile, generation 1 processor that's in a $999 ultrabook.
Click to expand...
I am not sure how much just 11.5 time faster is really "only", it is only 8 times faster than a small 4800U laptop chips as well, a 3990x is quite far from being twice a 3970x and so on, seem hard to fully scale pass a certain point on it.
 
Last edited:
Huge If. Depends on what the kernel does past 32 cores, if they go with unified memory, can they make big fast chips without going to chiplets?

Does a company that is allergic to VMs even wanna compete in that space?

Apparently PCIe 5 is due real soon and is gonna change the whole server spaces, with DDR5 too, so Zen4 Epyc is a lot more interesting than Zen3 which is mostly just inflicting trauma on Intel.
 
idiomatic said:
Huge If. Depends on what the kernel does past 32 cores, if they go with unified memory, can they make big fast chips without going to chiplets?

Does a company that is allergic to VMs even wanna compete in that space?

Apparently PCIe 5 is due real soon and is gonna change the whole server spaces, with DDR5 too, so Zen4 Epyc is a lot more interesting than Zen3 which is mostly just inflicting trauma on Intel.
Click to expand...
Yeah Intel's new Xeons have those according to leaks, they also have the networking, Storage, and supposedly GPU tech to take advantage of it. Also looks like Intel is finally able to pull off multi-dye chips so fun things are coming in 2021.
 
LukeTbk said:
I am not sure how much just 11.5 time faster is really only, it is only 8 times faster than a small 4800U laptop chips as well, a 3990x is quite far from being twice a 3970x and so on, seem hard to fully scale pass a certain point on it.
Click to expand...

Totally, I get what you're saying. But scaling is mostly dependent on how the software is written. For the main workload I do, scaling across cores is pretty linear. I upgraded from a 2990 to a 3990 and actually saw more than 2x the speed increase (as expected, since I gained in IPC and memory bandwidth as well). For some other programs I run, you are correct - the 3990x is not even close to twice as fast.

The only irritating thing about my main software is it won't take advantage of AMDs SMT. I can work around that by disabling it and clocking the CPU higher that it would boost to with SMT on, but it's still stupid that it only works on physical cores.
 
I work for a VMware shop and there is no enhanced vMotion support between Intel and AMD. So I would have to upgrade an entire cluster at a time. And of course our infrastructure budget got gutted this year due to COVID-related shortfalls. Going to have to keep stringing along these shitty Haswell-E systems even further.

The worst part is we are pigeonholed on old C7000 HP blade chassis. Have to forklift the whole environment to Synergy if we want to keep using blades. And HP sure pushes Intel hard right now from a product portfolio perspective. AMD still can't produce the sheer volume of CPUs that Intel can. And Intel still has that back channel into every OEM. You can get Epyc pizza boxes easily enough, but the depth of product offerings still isn't there.
 
NobleX13 said:
I work for a VMware shop and there is no enhanced vMotion support between Intel and AMD.
Click to expand...
shutdown vm, edit settings, options, under advanced->CPUID Mask, advanced, reset all to default

There's your migration path.
 
NobleX13 said:
I work for a VMware shop and there is no enhanced vMotion support between Intel and AMD. So I would have to upgrade an entire cluster at a time. And of course our infrastructure budget got gutted this year due to COVID-related shortfalls. Going to have to keep stringing along these shitty Haswell-E systems even further.

The worst part is we are pigeonholed on old C7000 HP blade chassis. Have to forklift the whole environment to Synergy if we want to keep using blades. And HP sure pushes Intel hard right now from a product portfolio perspective. AMD still can't produce the sheer volume of CPUs that Intel can. And Intel still has that back channel into every OEM. You can get Epyc pizza boxes easily enough, but the depth of product offerings still isn't there.
Click to expand...
I feel that pain, but hoping to get the last of my old E5’s out soon. But yeah it’s not going to be a fun job transitioning that last rack. And HP has to push the Intel, I would guess their AMD’s are back ordered and they are trying to move what they have in warehouse.
 
Mega6 said:
shutdown vm, edit settings, options, under advanced->CPUID Mask, advanced, reset all to default

There's your migration path.
Click to expand...
Last time I tried to migrate from Intel to AMD when you actually hit the migrate button to start the job it just flat errors out with a message along the lines of “Target platform is AMD, please select an Intel platform target”
 
Lakados said:
Last time I tried to migrate from Intel to AMD when you actually hit the migrate button to start the job it just flat errors out with a message along the lines of “Target platform is AMD, please select an Intel platform target
Click to expand...

Well what I typed works. Maybe you forgot the cpu mask.
 
NobleX13 said:
I work for a VMware shop and there is no enhanced vMotion support between Intel and AMD. So I would have to upgrade an entire cluster at a time. And of course our infrastructure budget got gutted this year due to COVID-related shortfalls. Going to have to keep stringing along these shitty Haswell-E systems even further.

The worst part is we are pigeonholed on old C7000 HP blade chassis. Have to forklift the whole environment to Synergy if we want to keep using blades. And HP sure pushes Intel hard right now from a product portfolio perspective. AMD still can't produce the sheer volume of CPUs that Intel can. And Intel still has that back channel into every OEM. You can get Epyc pizza boxes easily enough, but the depth of product offerings still isn't there.
Click to expand...
We are in the same boat.
That damn Intel vendor lock-in is bullshit, and I can't wait for that oligarchy to come to an end.
 
paradoxical said:
Totally, I get what you're saying. But scaling is mostly dependent on how the software is written. For the main workload I do, scaling across cores is pretty linear. I upgraded from a 2990 to a 3990 and actually saw more than 2x the speed increase (as expected, since I gained in IPC and memory bandwidth as well). For some other programs I run, you are correct - the 3990x is not even close to twice as fast.
Click to expand...
Was specially talking about cinebench R23 and had exactly what you are talking about in mind, with something that scale better that dividing by core to evaluate would have been really impressive, less so here.
 
Don't think Cinebench is a good benchmark as the core count goes up. The workload runs too fast.
 
Mega6 said:
Well what I typed works. Maybe you forgot the cpu mask.
Click to expand...
possibly, but in the end, we just waited 2 years so we could replace the whole stack. we weren't after a performance increase so much as a power load decrease. They Gotta keep their electricity bills down and increase those battery backups somehow, just means we get Milan instead of Rome.
 
LukeTbk said:
It is a big if thought right.

For example we could have used the exact same logic with zen 3, a 5800x score 15,245 on R23 and a 8 core 4800U is also doing much better than 88K / 16

Epyc is just 5.76 times faster than the 5800x while having 16 times more core, AMD is gonna smoke the Zen3 if the Zen3 can scale up.

The M1 does 1498 points in single core with the macbook pro and 7508 with all of them according to this:
https://www.notebookcheck.net/These...impressive-for-a-10-W-processor.504380.0.html

Scaling significantly less than Zen3 (I imagine that is because they have 4 performance core + 4 non full core ?) So maybe in fact they scale better and it would show if they would have 8 full performance core.


I am not sure how much just 11.5 time faster is really "only", it is only 8 times faster than a small 4800U laptop chips as well, a 3990x is quite far from being twice a 3970x and so on, seem hard to fully scale pass a certain point on it.
Click to expand...
lol I thought the same.
Competition is good but ARM still has some ways to go for consumers
 
Rockenrooster said:
probably because the M1 is in a consumer device that we can buy...
Click to expand...

So people is comparing an available mobile chip to an unreleased server chip because Ampere server chip isn't released still? This is a weird logic, isn't?

Both Milan and Altra Max will launch next year. This is the proper comparison: 64 core server vs 128 core server.
 
idiomatic said:
11.5 times faster with 16 times the cores? Apple is gonna smoke Zen3 if they can scale up. Intel's not even in the room any more.
Click to expand...

This is misleading. The M1 is an octocore, but HotHardware fails to mention it is a 4+4 configuration. Four performance cores plus four efficient cores. The efficient cores play a role similar to adding SMT to the x86 cores.

So it is not really 16 times the cores, but something like 32 times the cores. Or in terms of threads 8T vs 256 threads.
 
juanrga said:
This is misleading. The M1 is an octocore, but HotHardware fails to mention it is a 4+4 configuration. Four performance cores plus four efficient cores. The efficient cores play a role similar to adding SMT to the x86 cores.

So it is not really 16 times the cores, but something like 32 times the cores. Or in terms of threads 8T vs 256 threads.
Click to expand...
Its still cores though, not threads. Using the bigger cores would make the die even bigger and power usage would go up some.
adding 4 small cores take waaaaaay more die space than adding SMT.
4+4c vs 8c is more of a valid comparison than 4+4 vs 4c/8T any day...
 
I doubt the 64 core chip runs at the 4.8GHz my 5800X sustains while encoding video (it's dual 64 core chips, right?). Probably closer to 3GHz with that big of a chip.
 
The Apple M1 has system memory on die. I would wager that a good 20% of the performance boost comes from this. When you start having off-die memory, power goes up and performance goes down. This matters because the scores will look drastically different once your workload exceeds 16GB of memory, which is something that is easily done nowadays. While ARM has made big gains when you go core to core Intel and AMD are still ahead by quite a bit when it comes to overall performance.
 
You must log in or register to reply here.
Back
Top