Announcement

Collapse
No announcement yet.

16-Way AMD EPYC Cloud Benchmark Comparison: Amazon EC2 vs. SkySilk vs. Packet

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

  • 16-Way AMD EPYC Cloud Benchmark Comparison: Amazon EC2 vs. SkySilk vs. Packet

    Phoronix: 16-Way AMD EPYC Cloud Benchmark Comparison: Amazon EC2 vs. SkySilk vs. Packet

    With last week Amazon Web Services rolling out AMD EPYC cloud instances to EC2, I figured it would be an interesting time for a fresh benchmark look at how the AMD Linux cloud performance compares from some of the popular cloud providers. For this article are sixteen different instances benchmarked while looking at the raw performance as well as the value on each instance type relative to the benchmark performance and time consumed for the on-demand spot instancing. EPYC instances were tested from Amazon EC2, Packet.com, and SkySilk.

    Phoronix, Linux Hardware Reviews, Linux hardware benchmarks, Linux server benchmarks, Linux benchmarking, Desktop Linux, Linux performance, Open Source graphics, Linux How To, Ubuntu benchmarks, Ubuntu hardware, Phoronix Test Suite

  • #2
    Packet seems to be doing really well. VM vs dedicated machine, still price/performance seems great!

    Thanks for the tests Michael.

    Comment


    • #3
      BTW/FYI, 24hr pd x 365 days pa =$8760 pa. in theory per fully booked Packet instance..

      Comment


      • #4
        It would be much more interesting if we get benchmarks between EPYC and Intel instances with similar cost!

        Comment


        • #5
          Originally posted by kbumsik View Post
          It would be much more interesting if we get benchmarks between EPYC and Intel instances with similar cost!
          There will be an EPYC/Xeon comparison with the Packet.com instances on Wednesday~Thursday of this week published.
          Michael Larabel
          https://www.michaellarabel.com/

          Comment


          • #6
            I ran some mprime benchmarks on a m5a.24xlarge the day they were released. The most interesting thing to report was the NUMA configuration: each socket had 3 x 8 core chiplets. I was expecting 4 x 6 cores. I guess spreading out the thermal load wasn't a concern.

            My benchmarks showed a serious memory bottleneck at about 12 of the 24 cores loaded per die. Given the scaling I saw, I suspect each socket has only 3 memory channels, one per each chiplet. Generally it takes 4 Zen 1 cores to saturate 1 channel of DDR4 (or 2 Haswell+ cores per 1 channel of DDR4, as Haswell+ has double the AVX2 speed). Kind of a strange way to run Epyc, but I guess it's to better match the Intel offering, both in sizing (large, xlarge, etc) and in memory bandwidth.

            Comment


            • #7
              Thanks Michael for sharing the results. Packet.net was where I first got my taste of AMD EPYC usage on their hourly and spot pricing - spot pricing can drop that US$1.00/hr rate to US$0.20/hr for cheap testing if you time it right. But yes hourly billing is still expensive

              Comment


              • #8
                Originally posted by Michael View Post

                There will be an EPYC/Xeon comparison with the Packet.com instances on Wednesday~Thursday of this week published.
                Could you also try reaching out to german provider Hetzner(as they're monthly vs hourly with an upfront setup cost of 1 month making them a bit expensive to try out) to see if they'd be happy to make one of their systems available for these benchmarks? Would be nice to compare against Packet's bare metal offering, or is it just too different of a comparison/market?

                They offer a 128GB RAM AMD EPYC 7401P system for just 0.25 USD/hr effectively. Same processor as Packet is using but double the RAM at a quarter of the price, storage is 2x 960GB NVMe drives in RAID 1. https://www.hetzner.com/dedicated-rootserver/matrix-ax

                They also provide bare-metal systems with GPUs which I imagine would perform really well with some benchmarks against servers that don't offer that(and perform better than EC2 instances that do, but not against Googles TPUs for compute workloads).

                Comment

                Working...
                X