AWS EC2 instances based on AMD EPYC reviewed

AWS AMD EPYC review and benchmark 2018

Some days ago AMD announced that with AWS one more cloud provider is adopting its server processor line AMD EPYC. Those AMD based processors are custom built for AWS, based on the EPYC 7000 series and clocked at 2.5 GHz. They will be used throughout the EC2 product line for general purpose and memory optimized instances. I was eager to know how those new instances compare to existing Intel based EC2 instances: it's benchmark time!

In general I think it's great that AMD is on track again and that they seem to conquer some market share in the server world where Intel was omnipresent for the last few years. Competition is always good and Intel seems to be struggling, so AWS introducing AMD based EC2 instances is great for us users. With M5a and R6a, AWS introduced two new product families, which are around 10% cheaper than the Intel based M5 and R5 families.

I picked three Intel based plans and three AMD based plans for benchmarks and compared them with each other to see if not only the cost is lower but also if the performance is similar or even better on those AMD machines.

In particular I compared:

  • m5.large (Intel) to m5a.large (AMD)
  • m5.2xlarge (Intel) to m5a.2xlarge (AMD)
  • m5.4xlarge (Intel) to m5a.4xlarge (AMD)

AWS AMD EPYC CPU performance

To benchmark CPU performance I used a generic sysbench test to see how single and multi threaded workloads compare to each other. To have a more realistic scenario I also did some video transcoding with FFMPEG from a 4K video down to 1080p to see how those instances compare. You can download the video used for this benchmark here. Inside those instances the following CPU models where used:

  • Intel Xeon Platinum 8175M CPU with 2500.000 Mhz
  • AMD EPYC 7571 with 2199.868 MHz

What I think is strange is that the allocated AMD CPU only has 2.2 GHz, since in the AWS announcement they are talking about 2.5GHz, which would be similar to the Intel CPU those instances were spawned with.

Plan Threads Sysbench (MT) 4K to 1080p
m5.large (Intel) 2 15.5446s 89.556s
m5a.large (AMD) 2 17.2487s 87.879s
m5.2xlarge (Intel) 4 7.7994s 44.031s
m5a.2xlarge (AMD) 4 8.8486s 43.694s
m5.4xlarge (Intel) 8 3.8955s 22.637s
m5a.4xlarge (AMD) 8 4.9683s 25.619s

Additional benchmarks

If you've read my Amazon Lightsail review you know that those Lightsail instances are heavily throttled. I will do the same benchmarks with regular EC2 instances very soon which will also include those new AMD EPYC based instances. This is why I have skipped them in this post. If you are interested in and want to be the first to read those additional benchmarks I recommend subscribing to my newsletter.

Conclusion

The benchmarks show that those AMD EPYC threads are on average performing nearly identical to the Intel ones. From the first look, without testing multiple real scenarios, I think I would prefer those AMD instances, since they are 10 % cheaper without any noticable performance hit. Looking at the video transcoding benchmark it actually is faster on the smaller instances, so while getting cheaper instances they are also performing better sometimes.


Feedback and Suggestions

I always appreciate feedback and suggestions. You can contact me directly via niklas [at] karoly [dot] io.