AMD Rome EPYC Zen2 reviews

auntjemima
So, looks like they tested both 1p and 2p configuration of both processors. The 7742 easily beats the 8280 in many tests, even 1p vs 2p. In some it's close, and in some it loses handily. Really, you have to look at the results yourself, because they're all over the place. The important thing to remember is that each of the tests on phoronix is very specific to a particular task (except for a few), and unless you know what you're looking for they won't mean much to you.
 
Ok, so I rarely keep up with these launches, as I'm still on Haswell and happy. But I have a question.

I only read the Phoroix article, but it seems that the Intel 8280 chip is 28/56 and the AMD 7742 is 64/128? Am I reading that right?

If that's the case why is everyone surprised it is beating the Intel chip? At first I was unsure if my numbers were correct, but I checked the first page opening notes and it states the AND chip beats Intel considerably in both cores and threads.

Am I missing something?

Don't get me wrong, the power usage being fairly close on a chip with more than twice the cores and threads made me lol.
I guess its about how much $.
 
Ok, so I rarely keep up with these launches, as I'm still on Haswell and happy. But I have a question.

I only read the Phoroix article, but it seems that the Intel 8280 chip is 28/56 and the AMD 7742 is 64/128? Am I reading that right?

If that's the case why is everyone surprised it is beating the Intel chip? At first I was unsure if my numbers were correct, but I checked the first page opening notes and it states the AND chip beats Intel considerably in both cores and threads.

Am I missing something?

Don't get me wrong, the power usage being fairly close on a chip with more than twice the cores and threads made me lol.
"same core # VS same core #" is an arbitrary comparison. AMD just released a product that trashes Intel's best in performance in same server form factors. It destroys it in performance while using less power and a noticable reduction in TCO due to simpler cooling design and running cost requirements, at a significantly lower CPU price. Seriously, this sort of domination in almost every metric is practically unheard of. Nobody, outside some very specific situations, would turn down more than twice as many threads and a huge performance boost for much cheaper, just because, you know, "they have a better CPU in all regards that allows them to dominate in cores so of course it's faster".

If it was just a simple matter to put more cores in, Intel would have done it already.
 
Goes twice as fast, 40% cheaper.

Yet apparently this would be impossible for them to do to nVidia because business.

But still, imagine nVidia getting this colossally murdered. A card fully twice the speed of a 2080ti for the price of a 2070. That's how far up shit creek Intel is right now.
They focused on their CPU first, this has been stated for a while. They are making money now and can spend a bit on the GPU side again, hopefully we see good things in the future.
 
Ok, so I rarely keep up with these launches, as I'm still on Haswell and happy. But I have a question.

I only read the Phoroix article, but it seems that the Intel 8280 chip is 28/56 and the AMD 7742 is 64/128? Am I reading that right?

If that's the case why is everyone surprised it is beating the Intel chip? At first I was unsure if my numbers were correct, but I checked the first page opening notes and it states the AND chip beats Intel considerably in both cores and threads.

Am I missing something?

Don't get me wrong, the power usage being fairly close on a chip with more than twice the cores and threads made me lol.
Some of them are really confusing, they used Quad Intel i some benches, duals in others and singles in some... It's confusing.
auntjemima
So, looks like they tested both 1p and 2p configuration of both processors. The 7742 easily beats the 8280 in many tests, even 1p vs 2p. In some it's close, and in some it loses handily. Really, you have to look at the results yourself, because they're all over the place. The important thing to remember is that each of the tests on phoronix is very specific to a particular task (except for a few), and unless you know what you're looking for they won't mean much to you.
Also, careful reading some of those. As noted, the mkl-dnn and svt tests are both Intel written software. Also keep in mind, a single 7742 is a $7500 part, the Intel 8280 is a $13k part. If you compare single to single, AMD wins easily. If you compare and single vs Intel dual, it's much closer... But then your comparing a $9k system to a $30k system... And it still keeps up and wine ofte. It to mention has way lower energy usage, more pcid bandwidth (check some other reviews that did network testing with 100gbps, Intel has no bandwidth to compete due to low count of pcie lanes and stuck on pcie 3).
 
But relevant. I mean, if you can have 50 VMs running "ok" vs. 40 VMs running pretty darn fast. I know as a "lab" user, feeling "fast" is important to me in addition to scale.
ST performance != MT scaling where AMD does better again. Doesn't tell much bar clockspeed wins.
 
But relevant. I mean, if you can have 50 VMs running "ok" vs. 40 VMs running pretty darn fast. I know as a "lab" user, feeling "fast" is important to me in addition to scale.

I mean, it's good info to know, shows how close to the same speed a single core really is. In real world though, not very useful. I run a 12/24 at my house with 96gb of ram... I'm not very concerned with single core speed, I rarely have a single work load running and it's almost never a single threaded task if I do. Look at their context switch speeds and everything else, it can switch between a task just as fast as Intel and scales better (so, can utilize more cores more effectively). I would love to grab a single 64/128 for the house, but I don't think my wife would let me, lol. Something about bills and kids, although it's much more reasonable than the Intel equivalent.
 
I mean, it's good info to know, shows how close to the same speed a single core really is. In real world though, not very useful. I run a 12/24 at my house with 96gb of ram... I'm not very concerned with single core speed, I rarely have a single work load running and it's almost never a single threaded task if I do. Look at their context switch speeds and everything else, it can switch between a task just as fast as Intel and scales better (so, can utilize more cores more effectively). I would love to grab a single 64/128 for the house, but I don't think my wife would let me, lol. Something about bills and kids, although it's much more reasonable than the Intel equivalent.

But.... you see, I work where they decided to buy the very highest end (Intel) for our hypervisor hosts (speed and cores, favoring speed). And I can tell you it's very very difficult to go back.

With that said those 24 "slow" cores are cheaper. But they feel twice as slow (obviously an exaggeration).

So... when I chose my next lab build at home, I went with the faster cores vs. more cores. Maybe in a large scale scenario, I'd say well, I can run more VMs than you, but generally speaking my VMs twiddle their thumbs so much as is, that scenario (for me) is just not going to happen.
 
But.... you see, I work where they decided to buy the very highest end (Intel) for our hypervisor hosts (speed and cores, favoring speed). And I can tell you it's very very difficult to go back.

With that said those 24 "slow" cores are cheaper. But they feel twice as slow (obviously an exaggeration).

So... when I chose my next lab build at home, I went with the faster cores vs. more cores. Maybe in a large scale scenario, I'd say well, I can run more VMs than you, but generally speaking my VMs twiddle their thumbs so much as is, that scenario (for me) is just not going to happen.
Pretty sure that Rome would be just as fast or faster with less power usage...
 
But relevant. I mean, if you can have 50 VMs running "ok" vs. 40 VMs running pretty darn fast. I know as a "lab" user, feeling "fast" is important to me in addition to scale.
Single core performance is not that different. 10-20%. So it's more like a 100 VMs running fast vs 50 VMs running faster, IF the VMs run single threaded software only. If we disregard the whole I/O side of things where it doesn't look good for Intel, comparatively.
 
But.... you see, I work where they decided to buy the very highest end (Intel) for our hypervisor hosts (speed and cores, favoring speed). And I can tell you it's very very difficult to go back.

With that said those 24 "slow" cores are cheaper. But they feel twice as slow (obviously an exaggeration).

So... when I chose my next lab build at home, I went with the faster cores vs. more cores. Maybe in a large scale scenario, I'd say well, I can run more VMs than you, but generally speaking my VMs twiddle their thumbs so much as is, that scenario (for me) is just not going to happen.
But the single core speed in most tests was almost the same .. so your not giving up anything in worst case scenarios, but gaining a lot in the more normal use cases. So, you can have great single core and massive multi performance at lower prices. I understand if you just upgraded, it won't happen again for at least some time. My home server rarely breaks 10%, but sometimes I have all 24 threads @ 100%
 
Oh, and I would have a hard time convincing my company to switch to AMD as well due to government contracts. But, the server I use the most runs multiple VMs and databases and would be very happy for more cores (aka, less machines or more VMs with the same # of machines).
 
"same core # VS same core #" is an arbitrary comparison. AMD just released a product that trashes Intel's best in performance in same server form factors. It destroys it in performance while using less power and a noticable reduction in TCO due to simpler cooling design and running cost requirements...

Worth reposting. As if those investing in their server setup care about a "fair comparison".

This is a crushing blow to Intel.
 
Oh, and I would have a hard time convincing my company to switch to AMD as well due to government contracts. But, the server I use the most runs multiple VMs and databases and would be very happy for more cores (aka, less machines or more VMs with the same # of machines).

Too bad Microsoft just changed to core licensing instead of socket. :(
 
Worth reposting. As if those investing in their server setup care about a "fair comparison".

This is a crushing blow to Intel.

AMD knocked it out of the park. But the crushing blow has to happen in the purchasing departments.
 
AMD knocked it out of the park. But the crushing blow has to happen in the purchasing departments.

It's happening already, ie. in progress.

https://www.eweek.com/servers/amd-posing-major-challenge-to-intel-for-server-leadership

Jen Fraser, a senior director from Twitter, came on stage next and spoke to her company’s critical performance needs. One of their biggest problems is keeping energy use down, and the efficiency of this new second-generation Epyc processor is what attracted them to the product. Using this part, they were able to put 40% more cores into their data centers, using the same energy as what these parts replaced. Twitter plans to swap to Epyc 2 by year’s end.

The U.S. Air Force and Indiana University will both be early adapters of this new supercomputer technology. Indiana University indicated that it believes this technology will significantly improve the research-based advancements coming out and that its new Epyc 2-based supercomputer will be able to do in one second what prior technology took 28 days to accomplish. This will help the university continue to bring in the substantial amount of grant-based income ($750 million) that it currently receives.
 
^^ Exactly, nobody said it was going to happen overnight, but we could see big changes in 5-10 years with this type of cost savings improvements.
 
But the single core speed in most tests was almost the same .. so your not giving up anything in worst case scenarios, but gaining a lot in the more normal use cases. So, you can have great single core and massive multi performance at lower prices. I understand if you just upgraded, it won't happen again for at least some time. My home server rarely breaks 10%, but sometimes I have all 24 threads @ 100%

I was talking in generalities as to why the benchmark matters. You gave the result teeth... which again, emphasizes why it's an interesting benchmark.
 
AMD knocked it out of the park. But the crushing blow has to happen in the purchasing departments.

For me, that was helped by the recent announcements by HPe and Lenovo... Bueller? Bueller? (I mean, Dell? Dell?)

Looks like Dell "has them", but for whatever reason isn't exactly blowing trumpets.... Update: I hear they aren't quite ready to ship.

Update again: Perhaps some are "blowing trumpets" without the ability to ship :)
 
Last edited:
I was talking in generalities as to why the benchmark matters. You gave the result teeth... which again, emphasizes why it's an interesting benchmark.
I got you, just in my workloads (at home and work) I haven't seen much use in single thread performance on old or new hardware. If you have an all core speed of 2ghz and boost speed of 3.5ghz and you're using your server for more than one single threaded app, it doesn't matter what it *could* have done @ 3.5ghz if it's always (or mostly) running at all core speeds. If your server is idling most of the time and you have infrequent single core operations that must take place quickly... I guess I could see it, but seems silly to use a many core cpu for this use case.
 
Most computer twiddle thumbs... but when you need them, you like them to respond quickly. That's my point, and I'd argue it's the majority case.
 
Most computer twiddle thumbs... but when you need them, you like them to respond quickly. That's my point, and I'd argue it's the majority case.
Not in server world.
 
  • Like
Reactions: mjz_5
like this
Most computer twiddle thumbs... but when you need them, you like them to respond quickly. That's my point, and I'd argue it's the majority case.
I'm not sure, if that's the case, and you get an influx of connections at once, you need threads to handle them. If you only expect a single connection, what are you doing looking at 64 core system setups? I can't imagine a server feeling slow if it's barely doing any work unless grossly misconfigured. If it feels slow with a single thing happening, then it would be complete crap when loaded when the per core speeds come down.
 
Most computer twiddle thumbs... but when you need them, you like them to respond quickly. That's my point, and I'd argue it's the majority case.

Computers in homes yea. That's the opposite for corporate or dc. You want your servers working 24/7 otherwise you are losing money.
 
Computers in homes yea. That's the opposite for corporate or dc. You want your servers working 24/7 otherwise you are losing money.

I know that's the (popularized) goal. Have as much traffic on the freeway, bumper to bumper all day, all the time, because it's good... but you know, it's usually not good.
 
The only server we've ever had in any place that I worked where we wanted 100% utilization was a mainframe.

We're due for a datacenter refresh in q2 of next year and we're giving these a serious look. It's the first time we've considered AMD in the server space.

It does suck that MS changed their licensing to cores instead of sockets, at least VMWare and Veeam still license per socket. This could drop our maintenance costs on those products considerably.
 
I see a lot of planned OEM support for these but I'm not seeing much for orders. I wonder if AMD is hitting stock issues with the OEMs already?
 
I see a lot of planned OEM support for these but I'm not seeing much for orders. I wonder if AMD is hitting stock issues with the OEMs already?

Well, they just launched. I'd imagine it could be a little while before sales # and reports are showing up.
 
I know that's the (popularized) goal. Have as much traffic on the freeway, bumper to bumper all day, all the time, because it's good... but you know, it's usually not good.

This is a very bad analogy. A better one would be having a train running with the cars not full.
 
I see a lot of planned OEM support for these but I'm not seeing much for orders. I wonder if AMD is hitting stock issues with the OEMs already?
It launched 2 days ago, and some very large companies (Twitter, Google) announced they are buying... I'm going to guess sales are going ok. New product with lots of demand, I wouldn't doubt it'll take a little while for stocks to stabilize.
 
I know that's the (popularized) goal. Have as much traffic on the freeway, bumper to bumper all day, all the time, because it's good... but you know, it's usually not good.
The goal is to put just right # of cars to lanes so traffic doesn't build up. Your saying you want an 8 lane road with a single Lamborghini... Great, you can deliver one thing quickly, and that's it. If you choose the correct # or lanes for those vehicles, you can eliminate traffic and get a lot more done. If your packing a server till it's crawling, you either need a better server or more servers with load balancing. Just seems like a very uncommon use case that you're describing.
 
Well, they just launched. I'd imagine it could be a little while before sales # and reports are showing up.

I guess I'm used to Intel's releases where OEM's are ready for orders when Intel announces their server stuff.
 
Back
Top