Puget Systems print logo
Read this article at https://www.pugetsystems.com/guides/1041
Matt Bach (Senior Puget Labs Technician)

Why are we only using DDR4-2666 RAM with Threadripper?

Written on September 29, 2017 by Matt Bach

For years, Puget Systems has primarily provided Intel-based workstations. This isn't because Intel gives us a bunch of money or because we hate AMD, it is simply that throughout this time Intel was the best choice for our customers. With AMD's Threadripper CPUs, however, this is finally changing as there is now a solid reason to choose AMD over Intel in a number of situations.

One thing that might appear odd to some is that we are sticking with just DDR4-2666 RAM for our Threadripper workstations. After all, almost every hardware review of Threadripper uses DDR4-3200 (although some downclocked the memory to 2666Mhz) and AMD seems to make a big deal about how their CPUs can use higher frequency RAM, so why are we limiting our systems to lower speed RAM? Are we just trying to hold AMD back because we secretly prefer Intel? The reason why we made this decision really comes down to a four key points:

1) The official "Max System Memory Speed" specification for the current Threadripper CPUs is DDR4-2667 [source]. Why they decided to call it DDR4-2667 instead of DDR4-2666 like everyone else is unknown, but going outside of official specifications is something we try avoid doing unless there is an extremely compelling reason to do so.

2) Using RAM faster than the official specification voids AMD's warranty (as it is technically overclocking). To be fair, we haven't heard any reports of AMD refusing to warranty a CPU because of overclocking, but as a workstation manufacturer it is important for us to keep within the warranty terms whenever possible.

3) Higher frequency RAM is less stable. There are plenty of users running Threadripper at higher speeds than 2666MHz so this is  a point that some may disagree with. However, if you start to read around forums you will see that many of these users had to try multiple brands or models of memory to get one a configuration that is stable and they often have to tweak settings in the BIOS. Hardware enthusiasts building their own systems might be willing to take on this amount of extra work, but our customers typically do not want to have to mess around in the BIOS to get their computer stable. If we decided to use higher frequency RAM we would certainly do this work for our customers to ensure that any system leaving our door is rock solid, but things like overclocking often need to be fine tuned several times over the course of a system's life to stay stable.

4) There isn't a massive performance benefit to using higher frequency RAM. We have not tested this ourselves in great detail since the previous three points are enough by themselves to keep us from using higher frequency RAM, but the few reports we have seen and our own limited testing indicates at most a ~6% increase in performance with DDR4-3200 RAM over DDR4-2666. There has been testing with Ryzen 8 core CPUs that show a larger difference, but from what we have seen this does not translate to Threadripper. And while any increase in performance is nice, but in most cases this is not going to be a mind-blowing difference.

We know that many AMD fans are going to disagree with us on some of these points, but one thing we would encourage these readers to consider is this: If DDR4-3200 is actually stable and a good idea with Threadripper, why did AMD decide to only officially certify DDR4-2667? Even though the performance gains with DDR4-3200 are not massive, wouldn't having their CPUs perform 6% faster only help them to sell more processors? To our eyes, the only reason AMD would decide to take a 6% performance "hit" by limiting their official specification to DDR4-2667 is because they did not consider it to be stable.

In the end, what it boils down to is that our customers come to us because they need a computer that just works, and anything that increases the risk of bluescreens or other system instability is rarely worth a small increase in performance. If we were talking about twice the performance that would change things, but it doesn't take much to completely negate the benefit from any performance gain they would see with higher frequency RAM. A single system freeze that causes them to lose their last few hours of work or intermittent bluescreens that makes their system even a tiny bit unreliable is often more than enough. If you are building your own PC and are willing to take on the potential risks, go for it! But to us, this is one of those times where there is a clear difference between someone building their own PC who loves to tinker with hardware and users who simply need a workstation that lets them get their work done with minimal hassle and annoyance.

Looking for our Threadripper workstations?

As our systems are being sorted more and more by use-case rather than hardware specifications, it may not be clear where our Threadripper workstations are. At the time of this post, we have (or will soon to have) Threadripper configurations available for the following software packages:

Tags: Threadripper, RAM, DDR4-2666, DDR4-3200
Niko Nikolov

This is good info,but maybe it coud be better if you post some rendering charts with a example of a 1950x using 2666 vs lets say 2800mhz just so that we keep it stable.
Maybe a arnold test and a unreal engine light map bake test.


Posted on 2017-09-30 10:43:08

We could test the performance of ddr4-2666 compared to 2800, 3000, or even 3200 but the point we are trying to make is that any performance gain is really a moot point. If we saw a 2% performance gain with ddr4-2800, it would still be outside spec, void the warranty, and be less stable then ddr4-2666. Same thing if we somehow saw a 10% performance gain.

Keep in mind that our time is very much a finite resource and we have a pretty large backlog of critical testing that will actually impact our customers. So while it may be interesting, unless the results are way outside of every indication it wouldn't change our decision so it isn't the kind of testing that is a priority.

Posted on 2017-09-30 15:22:26
Steven Van

The article ignores the fact Threadripper system are going to be 32 GB 64 GB systems perhaps making 3200mhz 16gb sticks is costly. My

MICRON 64GB 4X16 MTA16ATF2G64AZ-2G6H1is certified at 2666 at 1.2 v I have it running at 3200 at 1.3 v couldn't be happier.

Posted on 2018-05-07 02:14:51
Jonathan Emms

Here my thoughts:
- Infinity Fabric between cores is linked to RAM frequency
- Different workloads will have different performance gains depending on workloads

It would be good to have further information on performance results on 3200 memory for each benchmark as a comparison, even with a disclaimer saying you don't sell or support workstations with 3200 memory.

Posted on 2018-06-20 02:01:21
gary gaisler

Thank you for the clear concise explanation for what was (to me anyway) a complicated question Despite the woulda coulda whydincha comments below your explanation is extremely sound and for someone building a system who is more of a novice it makes the understanding of the issue very clear.

Posted on 2019-01-20 17:19:36
Nathan Fitzer

This information is now outdated we know that faster ram speeds greatly benefit the new construction of ryzen which is in all of threadripper because infinity fabric is dependent on system memory. I also don't think that it's fair that you include the official memory speed for the AMD chips since Intel also lists the same official speed for theirs but you have no problem using XMP for your Intel builds. Intel considers anything over 2666 to be a warranty voiding overclock as well. One of my customers referred me to this article when I was building them a system you're limiting your sales and I guess sending them to me because you refused to use a Ryzen mem calculator.

Posted on 2019-01-31 08:12:58

How do I use this Ryzen Mem Calculator? I've ordered a pack of 4x8GB HyperX Predator 3000MHz running at 2933MHz due to ASRock X399 Taichi support. But I guess I was rewarded with a bad stick that cannot run at that speed so I tooked it off. Thus I ended up with 24GB @ 2933MHz single-channel. Wondering to calc a better solution to order, can you help me?

Posted on 2019-06-11 10:25:08