Can anybody calculate or give an estimate on the performance of the new Nvidia RTX 50-series GPU's, the 5070, 5070 ti, 5080, and 5090, in terms of Relative performance and Average time per WU?

I looked on the GPU page and no info for the 5090. 4090 vs 3090 is on average 35% faster. Using that as a guess the 5090 would take 125 sec.

Then I looked at 4060 Ti which has an average of 440.5 sec but mine consistently completes < 287 or almost a 4070 Ti. I wish they had four columns on the GPU page: linux, win 11, win 10 and "other".

We'll know soon enough what the average is for the 5090.

Then I looked at 4060 Ti which has an average of 440.5 sec but mine consistently completes < 287 or almost a 4070 Ti. I wish they had four columns on the GPU page: linux, win 11, win 10 and "other".

It also depends on how many parallel tasks per GPU you run. You should check "tasks per 24 hours" metric instead.

With effort, I did set my linux GPU to run multiple tasks but only gained 3% as it runs one task at 97%.

I also tried on a my win 11 box and was a failure as tasks took 2.5x to run. Net negative.

I agree that tasks per day would be useful assuming 100% participation in one project but I have not seen that provided.

Hi. Are there any tutorials for setting parallel tasks on the same GPU?

Probably but this is what I received from a friend (Keith Myers) on another project:

I'm new to linux and it's been a challenge but I like challenges.

Before you switch to mps, make sure your kernel size is as large as your card can handle in settings.

You need to be root, not just sudo, to run the nvidia_cuda_mps application. So you have to open a Terminal, get to root and then initiate the mps server.

I just have a text file called start_cuda_mps, marked executable, containing

export CUDA_VISIBLE_DEVICES=0,1
export CUDA_MPS_ACTIVE_THREAD_PERCENTAGE=70
export CUDA_MPS_PIPE_DIRECTORY=/tmp/nvidia-mps
export CUDA_MPS_LOG_DIRECTORY=/tmp/nvidia-log
nvidia-cuda-mps-control -d

Which I start in the terminal. You invoke it and it does not print anything to the terminal. You just exit the terminal but if you check your running processes once you start Boinc, you will see the nvidia-cuda-mps-server process running in the System Monitor and in the nvidia-smi task list.

The mps server really does improve the processing speed. On my Einstein O3AS tasks, it cuts the processing time in half.

You need to adjust the devices visible on your host depending on how many Nvidia cards are in the host. You can adjust the thread percentage depending on how many concurrent tasks are running on the card. 70% for two tasks is a good value and 40% for 4-5 tasks per card is reasonable.

You have to quite the mps server daemon as root also when you want to stop it running. I use this command to stop it.

echo quit | nvidia-cuda-mps-control