2. Install a graphics card for desktop in a PC?
Hi guys,
I know that outside of the professional sector, methods for connecting two graphics cards are dead and that's not what this is about!
I have a GPU with 16GB of vRAM that I'd like to use entirely for AI applications. It's currently in a trial run on Windows, but I'll switch to Linux.
At least under Windows, the desktop and a browser with a few tabs and enabled hardware acceleration alone take up 500MB of vRAM. The simple solution—if you don't want to do any work—is to turn off the monitor, of course.
But what about installing a small graphics card from the same manufacturer and connecting the monitor to it (except when gaming), so that the GPU can work in peace with the large RAM? If the motherboard and power supply allow it, it should be doable, right? How do you set this up correctly in Windows?
(1 votes)
Loading...
Basically feasible. The small card would actually take over all desktop tasks.
But one thing should be taken into account, you could get into trouble with the accelerator assignment. That your AI application always only wants to access the same small GPU.
In addition, you’ll cut your PCIe bus in the hälfe, which could be noticed with the CUDA performance. If the data transmission takes 2x so long.
Are PCIe Slots always sharing the Lanes? When the mainboard X470 chipset “2-Way SLI,CrossFireX” (here irrelevant), and:
has, then it is x16 slot only x8 if you use the x8 slot for the 2nd graphics card? Losing in the RAM feed would be of course… doof.
If the desktop card has only 8 lanes, no thing would be even if it can 16.
The CPU indicates the clock. The CPUs for Ottonomal consumers usually have between 20 and 28 PCIe Lanes. 4 go to the chipset, 16 are released for PCIe extension slots and, depending on the platform, further 4 – 8 should be kept free for NVMe SSDs.
You always have 16 slots for all GPUs you connect. 1 = x16 | 2 = x8 | 4 = x4 etc.
Exemptions are the Xeon and Threadripper/Epic CPUs. they also get 60 – 128 PCIe Lanes from the CPU and can also easily supply 4 GPUs with full x16
It is generally divided symmetrically. Always speak x8 at 2x GPUs.
If you really want to make AI seriously, then look at another platform that you want to get multi GPU for the bigger models sooner or later. Or don’t worry about the 500 MB now consumed in Windows.
Otherwise you can play a Linux without GUI later and make your entire AI with SSH
Or… eGPU via USB3 10GBit. That’s his own lanes, I suppose? Or the 2nd unused NVME slot?
Thank you. Then I guess the next question is if I find graphics cards that only need 4x *g* or can I trim this in the bios for the desktop GPU?
is feasible, a lot you don’t have to adjust, should simply live
You got a igpu? Then you don’t need a new one. Best regards
But unfortunately the 5900X has no :-/