I am planning to build a high end workstation or a server.
The primary use is for application development - learning and R&D for which I will need to install Database, Reporting software like power builder, python, web development tools, etc. In future, I may run AI, machine learning modules as part of training, there should be option to upgrade as well
Following configurations I have chosen for different setups.
Setup 1: Workstation
CPU - AMD Threadripper PRO 5975WX - 32C / 64T
GPU - Zotac RTX 4070 Ti
Motherboard ASUS Pro WS WRX80E-SAGE SE WIFI
RAM- Corsair Dominator RAM 32GB 3200MHz (8 X 16GB) (as there are 8 lanes)
Storage SSD - Samsung 990 Pro 2 TB M.2-2280 PCIe 4.0 X4 NVME Solid State Drive
Storage HDD - SATA 4 TB
CPU Cooler - Kraken X73 360mm All-In-One RGB CPU Liquid Cooler
Case - Cooler master Mid ATX
Setup 2: Server
CPU - AMD EPYC 73F3 - 16C / 32T
GPU - Zotac RTX 4070 Ti
Motherboard - Gigabyte MZ72-HB0 (Dual socket)
RAM- Corsair Dominator RAM 32GB 3200MHz (16 X 16GB) (as there are 16 lanes)
Storage SSD - Samsung 990 Pro 2 TB M.2-2280 PCIe 4.0 X4 NVME Solid State Drive
Storage HDD - SATA 4 TB
CPU Cooler - Kraken X73 360mm All-In-One RGB CPU Liquid Cooler (2 nos)
Case - Cooler master Mid ATX
I would like to know which setup would be good from performance point of view.
My understanding is that setup 2 would provide option for future upgrades.
I know that Threadripper PRO 5995WX (64C/128T) would provide better performance compared to setup1- Threadripper PRO 5975WX (32C / 64T) and setup2 - EPYC 73F3 (16C / 32T * 2). However for sake of uniformity I have selected both setup from Zen 3, supporting DDR4 memory.
I would like to know your views on which setup would be better in terms of performance, future upgrades - Setup1 or Setup 2
You’re trying to compare apples and pears: the TR system has 32c/64t while the EPYC system only has 16c/32t. That gives the TR system an unfair advantage in performance. So, swap out the 73F3 CPU with a 7551P (which is a 32c/64t CPU) for a fair(er) comparison.
Also, I’d suggest going for the Kioxia U.2 drives: better performance then the Samsung drives (which may or may not have a firmware issue, depending on which batch you buy from).
I would encourage you to choose ECC RDIMMs when it comes to your RAM on Threadripper Pro, especially as this is recommended by most board manufacturers; Even more so on Epyc as they don’t support regular unbuffered DIMMs at all.
Thanks for your prompt response. I
n setup2, I selected EPYC 73F3 (16C/32T) with dual socket motherboard, so effectively it will be (32 C/ 64T). So I guess comparison is apple to apple only. Both setups are Gen-3, DDR4 supported upto 3200MHz, PCIe 4.0
My understanding is that dual socket CPU for server machine, would provide more performance than single socket EPYC 7543P (32C/64T).
In one CPU, does the processes use all 32C/64T effectively, I have doubts on this.
You mentioned EPYC 7551P is 1st Gen CPU whereas EPYC 7543P is 3rd Gen CPU.
Thanks for suggesting Kioxia U.2 drives, Are you referring to SSD storage here? Why you opted for U.2 instead of NVMe M.2 or NVMe PCIe? Any specific reason
Hey, will chip in on the RAM. Use ECC RAM and not non-ECC ram. Your EPYC CPU x2 will not work and with Threadripper, if you’re going so far with ram, you better not have errors that can corrupt anything. It will be slower RAM but at least you will be sure it is safe.
I doubt the board will POST with 16x consumer DIMMs. There is a reason you use Registered for anything more than 2-4 DIMMs.
Bigger form factor allows for higher capacity and enterprise drives are usually more durable and have better write endurance than consumer drives. If the system has the lanes to support U.2, barely any reason not to except for maybe a boot drive. Prices have gone down significantly for NVMe storage. Micron 7400 8TB drive is <100€/TB, so basically cheaper than expensive consumer M.2 drives
You will loose half (!!) of your PCIe lanes from either CPU to connect it to the other one. So you have 2 heat sources (read: double rate on your electric bill!) but no additional PCIe lanes for connectivity. And said CPU interconnect (AMD calls it their Infinityfabric) adds latency to processes when it needs resources from the other CPU/RAM bank. The EPYC scheduler is actually very efficient, especially on Linux. But that call is yours.
On the NVMe drives: the enterprise grade ones have considerably better IOPS then consumer drives*, pretty important reading your use case.
*often double or even triple what a standard consumer NVMe drive would offer.
Got more info on PCIe channel usage on Dual socket CPU motherboard.
On EPYC side, I can go with 7543P (32C/64T) i.e. single socket CPU motherboard and zen 3 architecture for now.
However my doubt is that will the server use all the 32 cores be used effectively for running the processes? What is your view on this.
In comparison with Threadripper PRO which is primarily for HEDT, EPYC server configuration will provide more features on security aspect and robustness on server running for long hours.
That entirely depends on the software you’re running. As long as said software is multi-core aware and doesn’t have a hard-coded single-core thread programmed in, the scheduler will assign tasks to all cores/threads.
I vote threadripper pro, what your doing is more conducive to workstation, not server
like others have mentioned go for RDIMMs Samsung is great, crucial has horrible RMA support
you want either the Asrock Creator or the asus sage But it must be the second revision, the first has horrible memory support and lacks PBO on 5000 series chips
there is a difference in coolers being “compatible” and “made for”, compatible will not cut it and that’s what that cooler is
In case you still want an AIO watercooling solution go with one that covers the entire heatspreader. You can use an Alphacool Eisbaer Pro Aurora 360 for example, it will cover the entire CPU since it is made for HEDT and the 360mm radiator will have enough surface area to cool a Threadripper or EPYC processor.
One thing to mention on the motherboards: As a fresh owner of the V2 Asus WRX80 board, I think UNLESS you plan to use a lot of higher power PCIe devices, I would strongly consider the ASRock WRX80 Creator as well. From what I am reading they seem to be a lot faster with BIOS updates.
I talked to Asus and apparently they are working on better memory support and a new support list, but it seems ASRock is faster and already seems to have better support.
I am noticing some little things that could need more polish and the experience is not quite as rock solid as it was on my old Asus X399-A Prime; Sometimes the board takes 10-30 minutes on code F3 before suddenly passing to AA and initiating boot when starting from a USB drive, there are some PCIe errors; I haven’t used the ASrock board, but I hope someone with hands on for that one can comment on this being the same or better. That said, at least in Germany Asus’ support is super fast to answer and seems to be competent for the limited interaction I had with them in the past, but it could vary in your country.
Long story short: Consider both ASrock and Asus’ board.
