Announcement

Collapse
No announcement yet.

A Curious Look At Eight Core Server CPU Performance From Intel Xeon Haswell To AMD EPYC Rome

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • Paradigm Shifter
    replied
    What really draws me to the single-socket Epyc setups are things like that Asrock Rack boards are the 16 core or 32 core Epyc chips with the possibility of 1TB (or 2TB, if 256GB LRDIMMs can be found) in an ATX form factor. There is a certain level of resistance where I work to rackmounted systems, while something that fits on a desk barely gets blinked at even if it costs $20,000...

    Also, slightly older server kit has always been cheap - I picked up a fairly stuffed server a few years ago that came with dual Xeons, 96GB of RAM, redundant power supplies and four 3GB GTX580s for less than a single one of those 580's would have cost me new when launched. The only thing I needed to add was HDDs, and it had 6 hot-swap SAS bays. Used server kit is more expensive in the UK/Europe than the US, but you can still get some rather impressive deals.

    Leave a comment:


  • cjcox
    replied
    Ok, this isn't perfect. I don't have a golden Ubuntu to run on natively, so I ran this on kvm and tried to match. What it is, is 32 vcpu representing running on a 16c/32t dual processor E5-2667 v3. Another reason, not perfect, is that there were at least 3 other VMs running at the same time. One was a 32vcpu Windows 10. Anyway, there were other bits of services running here there and yon. But, I felt that perhaps, at least for people looking to "low ball" an Epyc system to replace an older 2 CPU Xeon Haswell, I thought it would be could to see something comparable. Some of the benchmarks didn't run (I might look into it, but since this wasn't "ideal" anyhow, I'll probably let this stand).

    Updated: HP Z840 2 x E5-2667v3 kvm Ubuntu (lightly to mid loaded)





    Last edited by cjcox; 03-30-2020, 11:22 AM. Reason: Updated results with more filled in. The dual E5-2667v3 gets 10 wins (mainly because of the extra cores)

    Leave a comment:


  • willmore
    replied
    Originally posted by kjujik View Post
    Nice article! What's worth adding is that these Haswell Xeons have pretty good perf/$ right now. I use 3950X at work, but at a home workstation/server, a dumpster dived (ebay) Xeon 2680v3/2690v4. Why? I compile a lot, and you can get these parts for as little as 140$. Combined with some X99 mainboard and ECC RAM, you can get yourself a solid workstation computer for < 400$ which is more than a price of a (new) low-end EPYC CPU!
    If you want used server hardware, that's pretty easy to come by. I see a lot of Gulftown era Xeon boxes with tons of memory for <$200.

    Leave a comment:


  • kjujik
    replied
    Nice article! What's worth adding is that these Haswell Xeons have pretty good perf/$ right now. I use 3950X at work, but at a home workstation/server, a dumpster dived (ebay) Xeon 2680v3/2690v4. Why? I compile a lot, and you can get these parts for as little as 140$. Combined with some X99 mainboard and ECC RAM, you can get yourself a solid workstation computer for < 400$ which is more than a price of a (new) low-end EPYC CPU!

    Leave a comment:


  • willmore
    replied
    Originally posted by FPScholten View Post
    So on my laptop (4c 8t Haswell 4700MQ) i get about similar performance with 1080p X265 encoding as the 8c 16t Haswell Xeon part in this test. (Yes I am using Ubuntu, but with custom built Xanmod kernel). When looking at HTOP when encoding I see that none off the CPU threads ever reaches 100%, most are around 85-90% only one thread occasionally hitting 100%. This has me thinking that it is not the processor but other parts of the system limiting the speed of this process (mainly io throughput I guess).
    Yeah, for me, x265 has trouble using more than 4 threads effectively. On my old i5-3750X, I got about 85% cpu utilization (quad core, no HT). It's slightly better on my R7-3700X, but there's at least half of the CPU going completely unused in x265. It's not a good benchmark for large servers *as is*. You can make it a better benchmark by converting it to a throughput test--run as many transcodes as it takes to saturate the CPU. You may even benefit from turning off multithreading in x265 and just running one instance per core. That's how it's going to be used on large servers anyway--they're going to try to do as much transcoding per machine as possible. The amount of time it takes to do any specific transcode isn't critical, just getting the most work done on the least hardware matters. Netflix and youtube have had blog posts about this in the past when they've said that they don't bother threading codecs as that only hurts throughput in their use cases.

    Leave a comment:


  • FPScholten
    replied
    So on my laptop (4c 8t Haswell 4700MQ) i get about similar performance with 1080p X265 encoding as the 8c 16t Haswell Xeon part in this test. (Yes I am using Ubuntu, but with custom built Xanmod kernel). When looking at HTOP when encoding I see that none off the CPU threads ever reaches 100%, most are around 85-90% only one thread occasionally hitting 100%. This has me thinking that it is not the processor but other parts of the system limiting the speed of this process (mainly io throughput I guess).

    Leave a comment:


  • intelfx
    replied
    Well, now that’s just seal clubbing.

    Leave a comment:


  • torsionbar28
    replied
    Originally posted by cjcox View Post
    So... the caveats.... while most Epyc servers out there are single socket, most (like 99%) of servers out there are using dual socketed haswell E5 v3's. (btw, there might a ton of v2 and even sandy bridge out there still, but usually dual socketed)

    So, better test would be to test a single Eypc against a dual E5. But yes, there are a limited number of dual Epyc's being deployed out there as well.

    I'm just trying to get you to a more likely scenario.
    We recently bought a bunch of EPYC servers, all were 32 core single socket. They actually replaced a bunch of older *four* socket servers. But in general, yes I agree that single EPYC is competitive with dual E5, in terms of price, performance, and product positioning, so that's the comparison that makes the most sense for benchmarks and reviews. Although if you just want to see how badly intel is getting their butt kicked, then yes by all means 1:1 lol.

    Leave a comment:


  • cjcox
    replied
    So... the caveats.... while most Epyc servers out there are single socket, most (like 99%) of servers out there are using dual socketed haswell E5 v3's. (btw, there might a ton of v2 and even sandy bridge out there still, but usually dual socketed)

    So, better test would be to test a single Eypc against a dual E5. But yes, there are a limited number of dual Epyc's being deployed out there as well.

    I'm just trying to get you to a more likely scenario.

    Leave a comment:


  • Nille
    replied
    does your numpy bench use the intel mkl?
    Last edited by Nille; 03-27-2020, 02:25 PM.

    Leave a comment:

Working...
X