AMD EPYC 48-core Review

Thanks to this zdnet article and, I was fortunate enough to receive a $250 trial of their bare metal server (Dell PowerEdge R6415) with the AMD EPYC CPU (24-core, 48 threads). c2.medium.x86 AMD EPYC intro

Persistent versus spot instance

If you want to use the bare metal dedicated server without interruption, you can pay $1.00/hr or ~$730.00/mo.

I opted for the spot instance (c2.medium.x86) and bid $0.20/hr. I experimented with spot instances at $0.10/hr and $0.15/hr. I failed to create spot instances at those hourly rates. It seems the floor for the AMD EPYC spot instance is $0.20/hr.

Not knowing the floor initially, I contacted tech support about setting up multiple servers. It seems they have to adjust your account so that you are entitled to more than one bare metal server at a time. Thanks to Samantha for this assist. The other issue was finding available servers. Thanks to Naji, she helped me locate servers outside the United States.

AMD EPYC bare metal server specs

Hardware specs:

Hardware Specs

  • 24 Physical Cores @ 2.2 GHz (1 x AMD EPYC 7401P)
    • 24 cores / 48 threads
  • 64 GB of ECC RAM
  • 960 GB of SSD (2 x 480 GB)
  • 20 Gbps Bonded Network (2 x MELLANOX 10GBPS W/ LACP)


OS Support:

Custom iPXECustom iPXE allows you to install a custom operating system manually or via automated kickstart.

 CentOS 7

 CoreOS – Alpha

 CoreOS – Beta

 CoreOS – Stable

 Debian 8

 Debian 9

 NixOS 18.03

 Scientific Linux 6

 Ubuntu 14.04 LTS

 Ubuntu 16.04 LTS

 Ubuntu 17.10

 Ubuntu 18.04 LTS spot instance vs Google preemptive instance

First off, when you create a Google Computing Engine instance, you can only select 24 cores at one data center. You can create an instance with more cores, but you have to ask for permission from Google first. For my quick comparison, I assumed Google gave permission and priced out an instance with 48 cores (48 threads). The cost for a Google Cloud Computing Engine 48 cores (48 threads) is $0.376/hr (730 hours/month, or 30.416 days)

If I compare the’s “spot” instance to Google Cloud’s preemptive instance, provides a greater value.

(Alternatively, if you create three 16 core/16 thread instances, it’d cost you slightly less at $0.36/hr ($0.12/hr x 3).

Google gives you high computing (preemptive is similar to spot instance) but I was surprised to find out that offers much more computing power at their $0.20/spot instance. I assumed Google Cloud would offer better computing to price ratio, but I was wrong. After all, Google has way more data centers and scale, yet blows them out of the water with the lower price. Google will introduce resource based pricing in Oct 2018, so we will see if the pricing gets any better.

Bottom line if you want the best value, is almost 50% less expensive than Google Cloud.

google computing engine $0.36/hr 730 hours $262.80 c2.medium.x86 $0.20/hr 730 hours $146.00

Persistent Storage Server

With packet’s bare metal servers, a spot instance is destroyed once another customer is willing to pay more per hour for that server. This means that whatever data you have on the server is destroyed. If you have to retain data, you have need to create and use a separate persistent storage server. I did not do this for my trial, but have some advanced guides on how to do this:

The advantage of Google Cloud preemptive instance is that your data is saved when the instance is stopped. You can always restart it with your data intact. Google gives you a max of 24 hours on a preemptive instance. Sometimes the instance stops in 5 minutes or actually goes the full 24 hours. I like’s spot instances because I’ve run them for many days without interruption. As long as you have persistent storage, it is easy to start another spot instance and retrieve your saved data from the storage server.

Using servers on the same network is certainly convenient, but if you want to save money, you can find alternative servers on on LowEndTalk/LowEndBox. And, of course, you can find great deals on bandwidth & storage servers in Europe such as OVH/KimSufi or Scaleway/ If you have a cloud service like Dropbox, you can also use free utilities like dbxfs to create a mount point on these spot instances. This is perfect for work units (BOINC).

Distributed Computing – BOINC – Rosetta@Home – World Community Grid

I used the AMD EPYC for crunching work units for cancer research (BOINC: Rosetta@Home (R@H) / World Community Grid (WCG). In return, I received BiblePay and ByteBall. BiblePay Team helps with cancer research and donates monthly to sponsor poor children (“orphans”) via non-profit charities.

In terms of performance AMD EPYC beat all the high end Intel CPUs in the pool that I participate in. Magnitude is the percentage of work units you contribute to the BiblePay team across several weeks. I was first in rankings with the XEON E5 behind. The other AMD EPYC spot instances were started after my first top ranking instance.

I found AMD EPYC CPU is awesome for tasks requiring parallel processing.

They’ve come a long way since 1998 with their AMD-K6-2 processor.

Intel has been mired in controversy recently with ill timed stock sales and CEO affairs. It certainly doesn’t help the company focus on innovation. AMD is on a tear with the AMD EPYC and Ryzen Threadripper. I wish AMD well and look forward to them giving Intel a run for their money!