Hi. I’m going to build a deep learning PC. Below are the components I chose for my build. I know this motherboard with 2x 3090 will run in pcie 4.0 x8/x8 mode which is equivalent to pcie 3.0 x16/16. In my experience, this won’t be a bottleneck. But maybe I’m wrong. Any advice and recommendations are welcome.
Type
Item
CPU
AMD Ryzen 9 5950X
CPU Cooler
Noctua NH-D15S chromax.black
Motherboard
Gigabyte X570S AORUS PRO AX
Memory
Kingston Renegade 128 GB (4 x 32 GB) DDR4-3200 CL16
Although, to be fair, I would only buy that heavy a machine if I could recoup the investment somehow, if it is just for learning either go with an Nvidia cloud solution or a single 3060. More is just waaaay overkill for what you want to do.
Although, to be fair, I would only buy that heavy a machine if I could recoup the investment somehow
^^this. years ago I spent 2999 on a Titan V and regret it since at the time I was only learning and thought if I got all the best I could get it would help me learn more side note, it didn’t I’m fortunate now that I have a job doing ML but if I were to plan a compute focused personal build id make sure I could make the money back first before investing.
I third this opinion - unless you have enough money that $8000 doesn’t mean anything to you, or this is a business expense from which you’ll be able to get that money back, you’re going to regret spending that much on this thing. Are you using this for business purposes or is it purely a personal project?
You would be much better served in my opinion by renting a cloud compute solution when you actually need the horsepower. You can get more compute power from one of them than you would ever even be able to run on the electrical circuit in your house and will keep scaling as the provider gets newer hardware.
To answer the question, I would get a server board, probably something from SuperMicro if you’re serious about this. If your application needs 128GB of memory, then I would recommend that memory be ECC as well, which puts you into a Xeon or Threadripper Pro platform. I wouldn’t trust any scientific application which uses that much memory to run for extended periods of time on a consumer ‘Gamer’ motherboard like that.
The 3090s still make very little sense. You get like 95% more CUDA cores with 4x 3080s, for pretty much the same price.
Of course, if you have infinite money to burn, that build looks sweet.
[edit]At a proper computer now. Let’s see, best bang for the buck:
GPU
Price
Price x2
Price x3
Price x4
Cuda
Cuda x2
Cuda x3
Cuda x4
3080 (10 GB)
$ 1039
$ 2078
$ 3117
$ 4156
8 704
17 408
26 112
34 816
3080 (12 GB)
$ 1239
$ 2478
$ 3717
$ 4956
8 960
17 920
26 889
35 840
3080 Ti (12 GB)
$ 1349
$ 2698
$ 4047
$ 5396
10 240
20 480
30 720
40 960
3090 (24 GB)
$ 1939
$ 3878
$ 5817
$ 7756
10 496
20 992
31 488
41 984
Best bang for the buck would be $ 4047 for three 3080 Tis as opposed to $3878 for two 3090. Even 3080s are more bang for the buck. The 3090s just don’t make enough sense to justify the high cost.
Yeah I get the feeling the OP might be in a similar spot, and if that’s true I wouldn’t buy in so big right off the bat. Chances are small networks will train just fine on what we call modest hardware these days, but I could be wrong, and they could be modeling what our friends at Los Alamos call “high energy particle physics simulations” or something like that.
Thank you all for your advice. Indeed, I analyzed my hardware needs over the past year and realized that I needed such a powerful build only twice, and at such moments I used cloud solutions. Here is my new build.
Type
Item
CPU
AMD Ryzen 9 5900X
CPU Cooler
Noctua NH-D15S chromax.black
Motherboard
Gigabyte X570S AORUS PRO AX
Memory
Kingston Renegade 64 GB (2 x 32 GB) DDR4-3200 CL16
Storage
Samsung 970 Evo Plus 1 TB M.2-2280 NVME
Storage
Samsung 970 EVO Plus 2 TB M.2-2280 NVME
Video Card
NVIDIA GeForce RTX 3090 24 GB Founders Edition
Video Card
GIGABYTE GeForce GT 1030
Case
Fractal Design Meshify 2 Compact
Power Supply
Corsair RMx (2021) 1000 W
What do you think about this build?
I chose a 3090 gpu, as I was faced with a memory limit quite often, and 24 GB is really necessary. The GT1030 card (will be installed in the third slot) is only needed for video output, since there may be delays when using the 3090 at full load.
Is the 1000W power supply redundant? According to calculations, it should be enough for 850 watts, but with a load of 90%.
All my components should fit in this case, I also plan to use all 6 (2x120 + 4x140) Noctua fans, but what about cooling? Does it make sense to take a larger case?
Is 1 TB ssd redundant for the OS? It may be worth saving and buying only 512 GB.
I also need a storage capacity for datasets and other files, about 4 TB. Can you recommend any solutions?