PCI x4 vs x16
19 Mar 2016 14:03:23 UTC
Topic 198515
(moderation:
Hey guys, is it worth putting a 512bit video card into a x4 PCI slot? The motherboard is new, has another x16 slot, which is filled. If it's put there, does it mean it will run at 1/4 out of 512bit? Or the difference is not that big?
PCI x4 vs x16
)
It depends, some apps are very heavy on PCI-E throughput, others not so.
Currently the difference is minor (<5% i guess) with the BRP4 and BRP6 apps. In the past it made a huge difference.
I suggest you try it, but of course power supply and cooling complications increase. Again you need more CPU resources to feed a second GPU.
Good luck, let us know before and after results.
Ha ha, thanks AgentB :) I
)
Ha ha, thanks AgentB :)
I was thinking about getting a sure answer from here, to see if it's worth buying the card or not. If I'm not wrong, BRP4 and 6 only use CPU, not GPU. CPU and power are not a problem, the power source is 1000W.
RE: If I'm not wrong, BRP4
)
The list of apps are here
BRP4 and 6 are GPU apps, the others are not.
All apps use some CPU resources, and the rule of thumb is to dedicate a core per GPU card to ensure the card is "fed" promptly.
RE: I was thinking about
)
I recently took the plunge into the dual card setup with this host doing two work units per card. BRP6 runtimes here.
One of those cards is in a x2 slot ;-)
AgentB, well, it's ok then I
)
AgentB, well, it's ok then I guess. It should work.
Gavin, thanks for sharing that! So how did you do it? First 1 card, to test it? Or did you start with both?
That machine originally
)
That machine originally housed an R9 390 which subsequently got moved to another host after recent the death of one of my 280X's...
The two 'Tonga' R9 380's were then just thrown in with the preference setting changed to allow two task per GPU. Whilst making the change my time was limited so I never got into the bios to make the second Pci-e run at slot x4.
I also have two 380's in this Linux host where the second slot is set at Pci-e x4... I still can't tell which task ran on which card!
Just note that neither of these hosts crunch on their respective CPU's. Whilst I reckon you should be good to go in adding a second card you may have to experiment with allocating CPU resources to find optimum performance if you want to also participate in the GW and GR searches.
Gav.
RE: Hey guys, is it worth
)
For others reading along: I assume you mean a "x16 sized" slot running at "x4" speed.
My recent experience with identical NVIDIA Fermi cards (470s) running the CUDA application for the Parkes work, two-at-a-time per GPU, one in a x16 speed slot and another in a x4 slot, is that if there is a difference it is very minor. Like...*very* minor. I will try this with a Maxwell soon, but have not done that yet.
Neither Fermi nor Kepler nor Maxwell cards seem to be negatively affected at x8 speed - but... the rest of any of my systems aren't exactly rockets. AMD CPUs and the attendant slow-down there, and the slower RAM, and the fact that the PCIe slots conform to PCIe 2 instead of PCIe3 standards... So, keep that in mind.
My most recent experience with ATI/AMD R9 270X cards is that there is no real difference running the OpenCL tasks for Parkes work, either.
I have no idea what might happen if you put a bleeding-edge card (a Nano or a 980Ti) in a x4 slot, but I would expect the OpenCL application to be affected more than the CUDA program.
BUT - all of my experience with this was either Parkes or Arecibo "Mock" and so anything other than those may be a completely different story.
OpenCL applications did demand more of the CPU, so I assume the PCIe bus as well.
I hope that's somewhat helpful.
RE: My recent experience
)
There is a long an detailed pair of BPR6 threads (sticky now) and a lot of graphics posted in those threads after a new release.
This post by archae86 shows my two identical Fermi GTX-460s (x4 and x16) and you can see that minor difference of maybe 1-2 percent slower in the "kink" on that graph.
Early days before app change from, the difference was 50% or more.
hth.