3Dwarrior
SOCIAL:
Filed under: Tech

First Look: Intel vs AMD EPYC AWS Cloud (IaaS) Benchmarks

AMD has made an epyc x86 comeback (pardon the pun) and the knowledge heart wars between the two chip producers is fleet heating up. Whereas gaming and user chips most often procure many of the clicking time, it’s the knowledge heart chips that power a bulk of the revenue and could well execute the most distinction whether an structure succeeds or fails. I will preserve my commentary to a naked minimum on this one and let the outcomes attain many of the speaking.

Background: The facts contained in this article become once equipped by an Intel spokesperson and all all assessments included can even be independantly verified by the exhaust of the same AWS conditions and configurations.

Efficiency comparison between Intel and AMD EPYC AWS conditions for cloud services (IaaS)

Amazon AWS EC2 offered AMD EPYC essentially based mostly conditions in November 2018. These conditions are powered by EPYC processors and have the same preference of virtual CPUs and reminiscence skill as the Intel conditions. The conditions can scale from 2 vCPUs the total attain up to ninety six vCPUs. Here’s the kicker though, on a generalized basis, AWS priced AMD conditions 10% more cost-effective than Intel conditions.

These conditions were the M5 (Intel) and M5a (AMD) fundamental reason conditions, R5 (Intel) and R5a (AMD) conditions. Currently we’ll be taking a have a study a enormous diversity of exhaust cases that for cloud infrastructure and accomplish an apples to apples comparison of those conditions (to the extent ability anyway). Utilize in thoughts, nonetheless, that while we could be definite AMD conditions are operating on the EPYC structure, Intel conditions is presumably now now not (Amazon would now not explain the structure) but since a majority of those assessments require AVX 512 we can steal that they are the most contemporary one.

Here is the total listing of workload kinds tested:

Int/FP workloads:

  • SPEC*rate2017_int_base 1 reproduction (Estimated)
  • SPEC*rate2017_fp_base 1 reproduction (Estimated)
  • SPEC*rate2017_int_base (Estimated)
  • SPEC*rate2017_fp_base (Estimated)

Memory bandwidth workload:

  • STREAM_OMP Triad

Web / Java workloads:

  • Server Aspect Java* 1 JVM
  • WordPress* PHP/HHVM

Database workloads:

  • HammerDB PostgreSQL
  • MongoDB

Excessive-Efficiency Compute workloads:

  • LAMMPS (Molecular Dynamics)
  • Excessive Perf. Linpack

So with none further ado, let’s soar into the benchmarks:


The first benchmark showdown is between the Sixteen vCPUs class of conditions. The benchmark is operating on Linux 7.5 the exhaust of the ICC compiler. Intel conditions have a dinky lead on web/java workloads as properly as a enormous lead in HPC. This is now now not namely handsome thinking in regards to the fact that EPYC has 2x 128-bit FMA while as Skylake and above have 2x 512-bit FMA – which provide them moderately the threshold in HPC initiatives.

I’d lift to declare two caveats here: a majority of those conditions, treasure the server facet ones, have now now not been NUMA-binded since that would require intensive tuning as properly as the fact that the exhaust of a numerous compiler (treasure GCC) can also consequence in improved efficiency. That mentioned, these are dear optimization routes and while they would slim the hole within the server-facet knowledge (numerous benchmarks treasure SPEC are already NUMA-binded) this can now now not be ample to beat the Intel counterparts.

A same describe emerges within the ninety six vCPU occasion where all once more, the Intel counterpart dominates the AMD facet of issues. The HPC distinction is all once more obvious however the reminiscence bandwidth has now been normalized.


At final, here is the efficiency per $ comparison the exhaust of AWS pricing as of twelfth January, 2019. Once extra here’s on a relative basis. On moderate, the Intel counterparts present higher rate from any place between 1.25x the total attain up to four.1x with HPC. What these assessments are attempting to declare here is that the Intel conditions provide both higher rate and absolute performances at some level of virtually about all cloud exhaust cases.

Once extra, you would possibly want to perchance well also most certainly procure the AMD numbers quite higher up by the exhaust of a numerous compiler and getting dedicated load optimizations in explain (corresponding to NUMA-binding) however the event ought to silent now now not be ample to overpower the underlying hardware differences (treasure MCM vs monolithic and the distinction in FMAs which is amazingly basic for HPC). On the premise of those numbers it looks that Intel does silent non-public the knowledge heart market – and that’s ok – since the educated/novice/video market is one where EPYC can in actual fact shine thinking in regards to the rate of Intel counterparts.

The submit First Look: Intel vs AMD EPYC AWS Cloud (IaaS) Benchmarks by Usman Pirzada regarded first on Wccftech.