Uh, to be honest, I’d avoid both. I urge you to adopt ECC memory for serious applications.
Go for a Xeon. If you like to stay with socket 2066, you can both adopt Xeon W-22XX (counterparts to 109XX), or W-21XX (Skylake). The latter can be found very cheapo in China, they ship worldwide.
Particularly cheap are the W-21x0B, originally made for iMac Pros. The W-2150B has 10 cores, the W-2170B does have 14 cores. Both do have 120w tdp, and produce less heat than your average 2066 cpu. The latter can be bought for ~400$.
Thanks again!
Will look into my options and share some RGB pics soon!
Unfortunately the great options suggested by you would take >1 month to reach my country (and will include some insane customs, Ex: on the Titan, I paid ~1000$ in customs!!) so I’d be forced to go with the 10-Gen processors
Understood. Go for the 14 cores then, if you can afford it. The 10 cores is also solid for configurations with <= 2 GPUs. I do have its Xeon counterpart on my windows daily driver, and I’m perfectly satisfied with it!
Can some of you expert recommends all the parts here https://pcpartpicker.com/list/ or similar website…? It will be nice for noobs like me who never build to have all the components so we can just order them and build ? =)
Yes 2 x 3090 =) My budget will be 5-6k and its mostly for deep learning / Kaggle … But maybe it will be good to have several brackets ? I am pretty sure there are a lot of people who will benefit from this =)
ECC Ram is not important for deep learning activities per se. It is important if you want to preserve an efficient and stable system in the long run.
Every time that a bit flip occurs in memory (and that’s more common than one thinks, see the relevant literature), and a file is open in memory, when the OS writes it back, it can write a corrupted file upon the filesystem. If that’s a system file, you will end up with problems.
I was a system administrator once, and I managed systems both with and without ECC. The latter always drifted towards instability, and required system reinstalls. The former, never.
Finally, ECC on the GPUs is only required if you do high precision numerical analysis with them (FP64 calculations). Indeed, the cards aimed at that (tesla, Gx100…) do have ECC.
That’s a very good idea. People have run extensive benchmarking about it, and it’s not even required to do awkward tricks like manual undervolting… Using nvidia-smi will suffice. It turns out that if you cap the maximum allowed power draw at 60%, you still enjoy some 90% of the card’s performance.
I do that regularly with my four GPUs (no space between them).
Often, we don’t appreciate the fact that manufacturers run their GPUs at their very limit in terms of power draw/heat produced/efficiency ratio, and that’s because even one single fps more in a game whatsoever is what really matters to them. For any serious use case that is inefficient and unadvisable.
Let’s take out 3k for the cards. This leaves us with some 2.5k for the rest.
If you don’t require sleep/hibernation, I’d go for a ROMED8-2T motherboard and the best Epyc you can buy with the rest of the money. Otherwise, go for a socket 3647 solution.
Thank you for the recommendations in this thread so far.
@balnazzar, in addition to your recommended ROMED8-2T motherboard and an Epyc,
is there anything to pay special attention to (with regards to 2x rtx 3090)?
Will a InWin R400N 4U Rackmount Server have sufficient space for the rtxs? any other recommendations?
Any additional airflow needed?
Will a Corsair MP600, Force Series fit under the rtxs on the board?
Also while looking at the Epyc 7282 I can only see that it as 64 lanes, not 128. I guess there a 7302 would be required?
Hard to say. It seems ok, but mind that the 3090 has unusual height, and the lock-bar of the case could interfere. I guess you should send an e-mail to in-win.
In my experience, space between cards counts more than case airflow, especially for non-blower card designs like the 3090 FE.
It will fit, but it could develop overheating issues, see post #9 in this thread. It would certainly overheat with designs like the previous generation Founder’s. The new Ampere’s heatsink avoids blowing how air orthogonally on the mainboard, but still, it will be better to install the ssd on a pcie add-on cards, since the ROMED8 has plenty of slots…
It has 128 lanes: EPYC 7282 - AMD - WikiChip. But for two 3090 it doesn’t really matter. They would work fine and without bottlenecks even with eight gen4 lanes each. Read it as: if you don’t care about ECC and registered ram (but you should), you can use a 3950X or any other processor with sixteen gen4 lanes.
Prior to go straight for two 3090, I would advise to wait a bit for knowing the street price of the new RTX A6000. Less headaches with parallelization, a lot less weight and torque exerted, standard card dimensions, no special requirements for the case, reasonable power draw, same amount of vram w.r.t. 2x3090.
Of course it would be interesting only if priced not much more than 3000$…
To add to the discussion, I’ve used a lambda labs PC at work, which has a 12 core i9-9920X processor and 4x 2080ti. For all my deep learning workloads, the 12 core processor is surprisingly sufficient. The avg load is about 6.0 on htop during training and I don’t think I have noticable I/O issues. An easy way to overcome I/O problems is to load your datasets into ramfs, so having 128GB RAM really comes in handy.
It’s a 44 lanes processor, gen3. It would be interesting to know which motherboard this machine uses: is there a PLX/PEX switch? If not, it confirms that even for a 4 gpu setup, 8 gen3 lanes are sufficient.
I noticed the same with an e5-2660v3 (10 cores, 40 lanes) and four gpus. Still, a beefier processor would come in handy for a plethora of tasks not directly related to gpu training (augmentation, preprocessing of big tabular data, etc…).
Agreed. The more, the better. And here is where rdimms win: much higher density per module, better price per gb.
Well its all plugged in and working! I am still trying to get the latest nvidia drivers (455) installed in ubuntu but everything on the windows side is working great. Maybe I can answer some questions if someone can lead me in the right direction to getting 455 installed in ubuntu. Right now cuda is not seeing the 3090s.
My bigger plan is to do a 6x 3090 FE setup using a dual case and water cooling. Basically I want to max out a consumer machine and not go server grade with epyc processors and a6000s or a100s
edit: Here is the current hardware:
CPU : AMD 3990x 64 core running at stock speed. Motherboard : ASUS ROG Zenith II Extreme Alpha TRX40 RAM : G.SKILL Trident Z Royal Series 256GB (8 x 32GB) DDR4 3200 GPUs : Dual 3090 FEs with +100 on core clock and +200 memory clock. 114% on power Boot drive : Samsung 2TB 970 EVO Plus NVMe PSU : CORSAIR AXi Series, AX1600i, 1600 Watt Case : Fractal Design Define 7 XL
Eventually I want to get a much bigger case, 2x 2000w power supplies, and 6x 3090 FEs. I want to mount the cards up and down vertically with the display ports at the bottom in a custom 3d printed enclosure. I would use a double case with the water cooling and motherboard on one side and the video cards on the other… Ekwb rads and blocks. I am thinking 4x 420 rads with two pumps and a large res. Each psu would be on a dedicated circuit with 4x 3090s on one psu and the rest of the system on the other psu.
