GPU power capping/limit

Hardware GPU
, ,
#1

Hello!

I just got a 3 year old media server and started to dust it out and reinstall the os on it. It has the following parts:
Asus Pro WS WRX80E-SAGE SE WIFI + AMD 3955WX + 4*8GB 3200 DDR4 + Silverstione sst-st1200-pts (1200W) + Nvidia quadro RTX 5000 + Nvidia Sync II + latest OS and driver (Win 11 pro 22h2 + nvidia 535.98)
Everything went great until I started to benchmark the system. The GPU performs really poorly like it scores 1/5 it’s normal score that others measured with the same GPU. After trying out everything I just noticed that in Furmark the GPU power is sitting at 20-24% att all time (my point was 380-400 and it should be at 2400 for the 1080p benchmark preset, also in notch builder my score was 3200 and should be around 21 000). I checked the nvidia-smi and the power limit is set to the maximum (230w) but it is usually at 45w when i check it under load.

I tried the following:

  • Replacing the gpu’s power cable from Y to two separate cables

  • Changing the nvidia “Power management mode” to “prefere maximum performance”, manually selecting physix processor to the RTX 5000, allowing the performance counters for all users

  • Disabling D.O.C.P, disabling above 4G decoding, disabling re-size BAR support, disabling sr-iov support, manually selecting pcie gen3 x16 mode for that slot

  • Disabling nvidia sync
    Everything else is configured for best performance in windows (power profile…etc)

Could anyone suggest me anything where to continue so that I could get all the performance out of the quadro card? (I don’t know if the card is deffective but after visual inspection I don’t see any clue)

1 Like
#2

I’m not an expert, but I wonder if it’s lacking cooling and is throttling down to 45W. I had this happen with a laptop I’d donated and it’d seen some abuse, bsd was creating fake load to keep temps down, and I could only conclude that the heatsink or fan was damaged.

1 Like
#3

Good answer above I agree you should check the cooling.
Let me look at my notes there are a few commands you could try in Nvidia-smi if you haven’t already.
1st: nvidia-smi -q -d SUPPORTED_CLOCKS
Then from the list of clocks find the highest memory and core clocks it allows.
2nd: example: nvidia-smi -ac 4001,2145
those are for a GTX 1660 yours will be different.
Manually setting those will still allow it to clock down when idle or should.
Watch your temps while under load with something like GPUZ.