Ryzen Threadripper Pro 3995wx Workstation Solver Build (2TB RAM)
Ryzen Threadripper Pro brings eight channel memory, as well official support for ECC memory to the Threadripper platform. Originally announced in the Summer of 2020, this platform only became available in Spring of 2021. Despite Threadripper Zen 3 (Theadripper Pro 5000x) likely to be announced some time in the second half of 2022, we still think this build will be of value to many looking for a very powerful solver workstation solution right now.
At the time of writing, the estimated cost of this build, configured with 2TB RAM, is $32,000.
AMD Ryzen Threadripper Pro 3995wx
This is a 64 core, 128 thread “enthusiast” workstation component. It bares some similarities to the AMD Ryzen 3990x, with it’s Zen 2 architecture, 280w TDP and 64 core count, but there are some noticeable differences. The notable differences for this build are a higher L3 cache of 256MB, which is relevant for processing data quickly in memory, 8 memory channels and support for up to 2TB RAM.
EVGA CLC 360 Liquid Cooling
AMD makes no recommendation as to what cooling should be used for the 3995wx on their website, but if we use the 3990x for reference, liquid cooling is the recommended cooling solution for this CPU. The EVGA CLC 360 will provide you with ample cooling performance.
If you prefer an air cooling solution, a compatible cooler that can also sufficiently cool the 3995wx under load is the Noctua NH-U14S TR4-SP3.
Supermicro is a California based IT company. They specialize in servers and high-end workstation solutions, among other products. Their M12SWA-TF motherboard is an e-ATX form factor server-workstation solution for Threadripper PRO that delivers 10G LAN and official support for 256GB RAM modules. This is required in order to run 2TB of RAM in this configuration. Note that there is no WiFi included on this motherboard. So you if would like WiFi, you will have to purchase a WiFi card separately. Regardless, we still recommend this motherboard at this time to users who are seeking a build with up to 2TB of RAM.
NEMIX RAM MEM-DR425MI-ER32 256GB DDR4 3200 (PC4 25600)
This RAM is listed on Supermicro’s website as compatible with their M12SWA-TF motherboard. For a 2TB configuration, you will need 8 modules in total. Always buy in kits if available. This reduces the risk of RAM failure/incompatibilities due to prior testing from the manufacturer for compatibility. For best performance, use at least as many RAM sticks as there are RAM channels available. sWRX8 has 8 channels of memory, so it is recommended that you use 8 sticks of RAM.
If you are uncertain whether you have selected the correct RAM for your build, we would recommend contacting NEMIX directly for assistance.
Seagate FireCuda 530 M.2 2280 4TB with Heatsink
PCIe 4 NVMe storage. For solver-workstation builds, we recommend to always have at least twice as much storage as RAM in the system. This is so that you can create a save file of any simulations that are currently running, that can be resumed for further calculation in the future. Additional storage needs may vary, purchase as required. For storing larger quantities of data, we would recommend using a NAS (network attached storage device).
NVIDIA GeForce GT 1030
Basic GPU that supports output to up to 2 monitors at 4k 60hz. Since this workload is not GPU dependant, we would recommend this basic graphics card. If you have needs that require you to go beyond this spec, say you are outputting to more than two monitors at 4k 60 or you need a higher performance graphics card for gaming purposes, we would recommend you choose a more powerful graphics card, such as a GTX 3060 for example.
Fractal Meshify 2 XL
This is a large and fairly straight forward to build in case with good air flow. If the design aesthetic is not to your liking, and you would like to select a different case, keep in mind that it should be compatible with e-ATX motherboards.
Seasonic PRIME TX-1000
Seasonic is considered the gold standard for desktop CPUs by many in the industry. This 1000w power supply provides ample power, with some headroom for future upgrades.
This is our current recommendation for a 3995wx solver workstation build with 2TB of RAM. We recognize that this build is very expensive, and mostly only of interest to a small subset of our readers, but a large part of the costs are due to the very expensive 256GB RAM modules. If you do not require 2TB of RAM, a 512GB RAM kit, for example, can be had for approx. $20,000 less, bringing down the total cost to about $12,000.
Please leave a comment below if you liked this build or you have any questions. Also let us know what other builds you might be interested in seeing in the future.
Disclaimer: We may receive a commission if you use one of our links to purchase any products on a partner site.
Any recomendations for builder or buying parts for this. How hard is to build such system DIY versus regular PC?
This is still a desktop computer, so it shouldn’t be significantly harder to build than a regular PC. The main difference is that the case design is larger, as it needs to support the E-ATX form factor.
As far as buying parts, all the parts are linked to if you click on the images of the parts. I don’t know what region you are in, but if you go to Google and type “3995wx workstation” and then go to the ‘Shopping’ tab, you will see pre-built machines for purchase. Those suppliers may also be able to customize a build for you.
It would be really cool if you could publish a hardware database of a standardised solve. It would be really interesting to know what kind of bang for your buck you get for r9 5950x/3995x/3990x/3970x/i9-12900k/rented servers/EPYC options. It would also be useful to quantify the benefits of ddr5 vs ddr4. I’m currently building a 3970x workstation and I’m toying with the idea of a 12900k ddr5 workstation for smaller (<128gb) solves.
Thank you for your comment Joe. I agree with your sentiment. Regarding ddr5 vs ddr4 on the 12900k, I think ddr5 will be the way to go moving forward, but one issue is just the availability and cost of high performance ddr5 kits at this time. A 12900k build with ddr5 post is planned, but at launch there wasn’t a clear benefit to using ddr5 over ddr4 due to ddr4 kits being relatively higher performing and so comparable to ddr5 kits. I would still build 12900k with ddr5, but I expect ddr5 performance to improve over time as better kits become available.
I was taking a closer look at the 12900k and it seems to me to be a poor choice for a solver build. The edge it has over the AMD chips seem to be single core performance and a low power draw when idle. Neither of those factors are relevant for a solver machine. It might have a small edge in absolute multicore performance but that comes at a cost of a much higher power draw. I think I’d rather wait for AMD to release their DDR5 compatible chips/boards. Hopefully DDR5 RAM prices will be more realisitc at that time.
I think that’s reasonable.
How important is RAM speed? In addition, how much of a performance penalty does Broadwell vs the Threadripper 3xxx? Thinking of getting a old server like a DL580 G9 with 4x 8894v4 – this gives 96 cores as DL580 G9 / 4x 8894v4 will set you back around $2500 (RAM extra) – would that be more economical and yield the same performance?
RAM speeds matter since data remains in RAM whilst solving. So the faster data can be access in RAM, the faster the solution can solve. Another issue is that the overhead from syncing a quad socket CPU configuration leads to a performance penalty.