AMD EPYC 7000 Series Specs and Performance Leaked

Megalith

24-bit/48kHz
Staff member
Joined
Aug 20, 2006
Messages
13,000
VideoCardz has published performance numbers for AMD's upcoming server processors ahead of the embargo date: some of the specs gleaned from the press materials include 32 high-performance "Zen" cores, 8 DDR4 channels per CPU, up to 2TB memory per CPU, and 128 PCIe lanes. Hit the link for further details and how they compare to the Xeon line.

The EPYC 7000 series processors feature 128 PCIe lanes and 8-channel DDR4 support (up to 2666 MHz). Some parts are listed with two TDP values (I’m not sure why). EPYC CPUs are available with up 32 cores. The cheapest part should be available [at over] 400 USD, and the most powerful EPYC 7601 processor will be sold for around 4000 USD. The highest clock speed is 3.2 GHz in turbo mode.[/URL]
 
Last edited:
NICE.

Now, show me this offered from the big guys. Cisco, Lenovo, Dell, HP. Show me blades and pizza boxes. Show me these parts serving as the guts for SANs and NASs.
 
Were they under any form of NDA? Still kinda crappy of them though.
 
Unless there is something very wrong with these chips there is no way guys like AWS aren't going to buy heaps of them.

The 8C/16T 2.1-2.9Ghz chip going for somewhat over $400 is a heck of a deal. The top end $4K chips are going for higher than I thought but would still offer good value vs. the competing top end Intel Xeon's so its not too terrible. Clockspeeds could be higher of course.
 
Lots of money will be saved at datacenters on per socket licensing alone.

I'd like to see more numbers. VMWare will need to get their Ryzen shit together before any major OEM will pick these processors up.
 
Were they under any form of NDA? Still kinda crappy of them though.

I don't think so. I say that on how it's worded: "We got our hands on EPYC 7000 series press deck. The NDA on this material ends on June 20th." To me this sounds like some 3rd party gave it to them. I could be wrong™.
 
Give me a 16C/32T CPU and I will have my workstation for the next 2 years. Unless for some UNGODLY reason we will get PCI Xpress 5.0 when it's supposed to come out and we can swap CPU's out.
 
  • Like
Reactions: N4CR
like this
The base clocks are too low. Let's say I'm looking at a 8 or 10 core Xeon that has a 3.3GHz and 3.2GHz respectively. That would kill the AMD Epyc 8-core @ 2.1GHz
 
Don't worry, the big players like MS and Oracle will just change up their licensing models to make more money like they did when virtualization came along.

They already did. Now the Standard Windows Server license only goes up to 8-cores.
 
Don't worry, the big players like MS and Oracle will just change up their licensing models to make more money like they did when virtualization came along.


MS already has. You pay by the core for their 2016 licenses, not by the socket. You even pay per core you assign to a SQL server VM.

VMWare and Veeam are still per socket though. So's pricing for XenServer's support (the software is free).
 
The base clocks are too low. Let's say I'm looking at a 8 or 10 core Xeon that has a 3.3GHz and 3.2GHz respectively. That would kill the AMD Epyc 8-core @ 2.1GHz

That's a bit of an awkward comparison. The Epyc is going to have 128 PCI-E lanes plus 8 DDR channels (and still probably costs much less than said Xeon). An 8 core Threadripper would be a more direct comparison in terms of overall performance.
 
  • Like
Reactions: N4CR
like this
The base clocks are too low. Let's say I'm looking at a 8 or 10 core Xeon that has a 3.3GHz and 3.2GHz respectively. That would kill the AMD Epyc 8-core @ 2.1GHz
Skylake Xeons are also going to have fairly low clocks. These chips are all about heaps of cores not clockspeed.

If you want peak clockspeed and lots of cores buy either a Threadripper or HEDT Intel and OC it.
 
That's a bit of an awkward comparison. The Epyc is going to have 128 PCI-E lanes plus 8 DDR channels (and still probably costs much less than said Xeon). An 8 core Threadripper would be a more direct comparison in terms of overall performance.

Servers. We are talking about servers. They are not putting Threadripper in servers and I can't buy a faster 8-core Epyc processors even if I wanted.
 
Skylake Xeons are also going to have fairly low clocks. These chips are all about heaps of cores not clockspeed.

If you want peak clockspeed and lots of cores buy either a Threadripper or HEDT Intel and OC it.

Tell me if I wanted the fastest 8-Core dual-processor rack server available which one should I get?

E5-2667 v4
or
Epyc 7251

ANYWAY, the top end Threadripper is the Epyc 7351P. I hope that's not the clock for that Threadripper. And if that is not the clock for Threadripper 16C and it's much faster clocks, why would I buy that Epyc 7351P when I can get the Threadripper...
 
Tell me if I wanted the fastest 8-Core dual-processor rack server available which one should I get?

E5-2667 v4
or
Epyc 7251

ANYWAY, the top end Threadripper is the Epyc 7351P. I hope that's not the clock for that Threadripper. And if that is not the clock for Threadripper 16C and it's much faster clocks, why would I buy that Epyc 7351P when I can get the Threadripper...

The Epyc 7351P will be much easier to upgrade down the line - the 7351P is at the bottom of the EPYC stack whereas the 16C Threadripper is at the top of its product stack. I suspect the TR CPU will also have a higher TDP than the 7351P's 155W to reach its higher clocks? (Maybe?)
 
Servers. We are talking about servers. They are not putting Threadripper in servers and I can't buy a faster 8-core Epyc processors even if I wanted.
There were Dell servers you could get with i3 and i5 processors instead of full out Xeons. Single socket of course, but it's not unheard of for a server to be sold using what is ostensibly a desktop part.
 
I would love to see them try this day and age, since stuff has a tendency to get "leaked" more often. But their fanbois will make sure to come out in hordes to defend them.


Hopefully the playing field remains even in that respect, if amd had seen that trial through to the end they could have came away with much more than they got.
 
Well, assuming Intel don't resort to their old tactics.
Intel can't bully guys like AWS and if Intel wants to compete on price I think AMD will actually have more room than Intel this time around since Intel has to support its fabs.
 
There were Dell servers you could get with i3 and i5 processors instead of full out Xeons.
Dell calling them servers doesn't necessarily mean they're actually server grade hardware though.

You can actually use almost any PC as a server depending on your needs but the relability won't be there.
 
Tell me if I wanted the fastest 8-Core dual-processor rack server available which one should I get?
The E5-2667v4 will probably be faster but cost lots more. The E5-2667v4 is a ~$2,000 chip while the Epyc 7251 is a ~>$400 chip. The Epyc 7251 should also support more RAM and have better I/O options too. Depending on your work load it may very well be only a bit slower than the E5-2667v4. Saving ~$1,500 per socket is also a fairly big deal too.

ANYWAY, the top end Threadripper is the Epyc 7351P. I hope that's not the clock for that Threadripper. And if that is not the clock for Threadripper 16C and it's much faster clocks, why would I buy that Epyc 7351P when I can get the Threadripper...
Look at the top clocks for the 24C/48T+ Skylake Xeons, they aren't all that good either. If you want lots of cores and sub 200W TDP's something has to give.
 
Tell me if I wanted the fastest 8-Core dual-processor rack server available which one should I get?

E5-2667 v4
or
Epyc 7251

ANYWAY, the top end Threadripper is the Epyc 7351P. I hope that's not the clock for that Threadripper. And if that is not the clock for Threadripper 16C and it's much faster clocks, why would I buy that Epyc 7351P when I can get the Threadripper...

I thought we we're talking about Servers :rolleyes:.

To answer your question, ECC RAM, more RAM channels, more PCI-E lanes, and other server-specific features of Epyc versus Threadripper. It really depends on the use case.
 
Skylake Xeons are also going to have fairly low clocks. These chips are all about heaps of cores not clockspeed.

If you want peak clockspeed and lots of cores buy either a Threadripper or HEDT Intel and OC it.

The 8180 is 2.5Ghz base, 3.2Ghz all cores turbo and 3.8Ghz max turbo on 28 cores for example. Its quite a huge difference. Not to mention single mesh without scaling penalties.

8180.png
 
The 8180 is 2.5Ghz base, 3.2Ghz all cores turbo and 3.8Ghz max turbo on 28 cores for example. Its quite a huge difference. Not to mention single mesh without scaling penalties.
Nah, except for the 3.8Ghz max turbo for a few cores, you're only looking at a difference of a few hundred Mhz for the most part vs a Epyc 7451 or a Epyc 7601. Yes the core #'s are different but AMD doesn't have a 28C/56T part. The closes AMD parts have 24C/48T or 32C/64T.

That Skylake Xeon 8180, which is the very top end of the new Xeon's many of the other Skylake Xeon's have low 2Ghz-ish clocks or less than 2Ghz clocks too, is supposed to cost over $12,000 vs the $4,000 Epyc 7601 too which is something you have to consider as well.

Also every sort of inter core bus/network/mesh/whatever marketing BS name you want to use is going to run into scaling penalties eventually. Nothing scales perfectly. Better than the old ring bus? Yeah I can believe that. Better than AMD's new IF bus? That remains to be seen.
 
Nah, except for the 3.8Ghz max turbo for a few cores, you're only looking at a difference of a few hundred Mhz for the most part vs a Epyc 7451 or a Epyc 7601. Yes the core #'s are different but AMD doesn't have a 28C/56T part. The closes AMD parts have 24C/48T or 32C/64T.

That Skylake Xeon 8180, which is the very top end of the new Xeon's, is supposed to cost over $12,000 vs the $4,000 Epyc 7601 too which is something you have to consider as well.

Also every sort of inter core bus/network/mesh/whatever marketing BS name you want to use is going to run into scaling penalties eventually. Nothing scales perfectly. Better than the old ring bus? Yeah I can believe that. Better than AMD's new IF bus? That remains to be seen.

Supposed? ;)

Infinity fabric isn't anything magical. In reality a single 7601 is more or less equal to an 8S system with full mesh. So yes, its a very crude setup with the limitations it brings.

Infinity fabric is equal to Omnipath, QPI, Hypertransport etc.
 
There were Dell servers you could get with i3 and i5 processors instead of full out Xeons. Single socket of course, but it's not unheard of for a server to be sold using what is ostensibly a desktop part.

Yup they used to even sell SB servers with P4s back in the days when Athlon Xp was dicking even the northwood p4s.
Our favourite intel share price shills must be having a wonderful day today.

Unless there is something very wrong with these chips there is no way guys like AWS aren't going to buy heaps of them.
The 8C/16T 2.1-2.9Ghz chip going for somewhat over $400 is a heck of a deal. The top end $4K chips are going for higher than I thought but would still offer good value vs. the competing top end Intel Xeon's so its not too terrible. Clockspeeds could be higher of course.

Those are the dual socket chips (4k) the single socket versions are far cheaper and 200mhz slower.
I am praying someone makes a TR style/HEDT orientated Epyc board which can OC. Less ram slots (e.g. 'only' 8) etc. The sockets are almost identical anyway.. 32 cores at 3.8GHz would be monstrous.
 
Yes that would be impressive but pricewise you're still looking at a $2,000 chip for the 1S only 7551P Epyc. I don't think there'll be much of a market for something like that unfortunately. That is part of the reason Intel's new HEDT line is getting the cold shoulder by many enthusiasts and there is more excitement for TR, the bang vs. buck ratio matters more these days.
 
  • Like
Reactions: N4CR
like this
Yes that would be impressive but pricewise you're still looking at a $2,000 chip for the 1S only 7551P Epyc. I don't think there'll be much of a market for something like that unfortunately. That is part of the reason Intel's new HEDT line is getting the cold shoulder by many enthusiasts and there is more excitement for TR, the bang vs. buck ratio matters more these days.

Very good point. In reality I'll have to settle for TR it's expensive enough as it is. But if they did do some super uber enthusiast tier board, it would make for awesome e-peen/WS builds. Most of my use is WS and some gaming. Had to do some more video editing this week on the damn temp i3 though.. FML. Nearly ready to just pick up a damn Zen board/cpu/ram... but i'll resist and wait it out for TR. Especially if the socket is as they say, compatible with 7nm next year. 32C 7nm HEDT yes please. Cheaper way to do it than Epyc now.

If these bad boys are unlocked, goodbye Intel.
You're just an AMD fanboy ;P
 
Supposed? ;)

Infinity fabric isn't anything magical. In reality a single 7601 is more or less equal to an 8S system with full mesh. So yes, its a very crude setup with the limitations it brings.

Infinity fabric is equal to Omnipath, QPI, Hypertransport etc.

Crude by what measure?
 
Crude by what measure?

By the interconnect ability and bandwidth. Just as seen on Ryzen between CCXes. Now scale up, take a 2S system with 8 dies in total. Each die got 2 GMI links and tell me how many hops and shared bandwidth links you need instead of direct.
 
What exactly is crude about the "interconnect ability" and/or bandwidth? How in the heck does bandwidth alone tell you anything about a given bus' complexity?! Especially since its supposed to scale from 10's to hundreds of GB/s in bandwidth depending on how they implement it.

And inter-CCX bandwidth has nothing to do with inter package or inter CPU die IF bus bandwidth either. Different application = different implementation = different performance with IF.

Bear in mind that there is virtually no detailed public information on the IF bus still at this point. We don't even know the latency at this point and that is every bit as important as bandwidth if not more so.

We also know very little about Intel's new inter CPU bus too. What little information we do have suggests its a evolution of the ring bus ironically. Which isn't a bad thing at all but does throw serious shade on any sort of commentary suggesting its somehow more advanced and/or capable of IF.
 
Until VMWare fixes all their Zen problems, these massive core/lane counts mean nothing.

It is astounding to me how badly AMD supported their partners in this launch.
 
Back
Top