Few days ago I had the chance to try NVIDIA RTX 4090 Founders Edition in Einstein@Home and oh boy it literally chewed up those Gamma ray pulsar Work Units in a minute. I ran it for 1 hour and 30 minutes and Einstein@Home granted me 155K points so far and it keeps increasing every day.
Interestingly it only ran Gamma ray work units and nothing else. And work units seemed frozen at 87% and then get completed...
I think it's all about to change. The new 4000L series of tasks may have a much shorter follow-up stage, based on past behaviour. Overall, they may take a little longer to crunch since the 0-~90% stage for 4000L series used to take measurably longer.
work units seemed frozen at 87% and then get completed...
There's basically two parts to a work unit. Firstly there is the bulk of the work in producing a ( fast ) Fourier Transform from a given time series of measurements. Secondly the ( candidate ) signals - one of which may be a discovery - from the first stage must be ranked and sorted into a list to be returned to the project. So the % measure is certainly not a time indicator, as you have discovered. In fact I don't know/remember which part of the code does emit the % amount, nor how it calculates that. Does anyone know ?
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
I also noticed only half of the GPU was being utilized and it was fairly cool (50ish celcius degrees)
That means you could try running more than a single task at a time and see if two task running at the same time take less time that each task running individually, if so and you still have headroom left you can try 3 or even more tasks at a time. Remember though that each gpu task needs cpu time and just leaving a single cpu core free to do that may not be enough to run at full speed.
The a) set-up, b) FFT, c) find signal candidates (abc-abc...) loop is done continuously in most Einstein task types and on some other projects (setiathome) too. i.e. the MAIN-loop.
Sometimes the after processing is done on CPU after the MAIN-loop and then you see a halt in progress percentage. That slow-down could be faked to seem like progress, but it actually represents the time it takes to do a similar amount of (flops) work.
With Nvidia and Cuda Linux executables you may be able to use cuda-mps-server to run multiple program executables in a (super) parallel manner to overcome the pre- and post processing halt. See the NVIDIA documentation.
Question- I just updated the NVIDIA drivers and I received an odd message in BOINC that I am not sure I understand. I am not sure if it is related to the new drivers or the BOINC build I also recently updated.
11/29/2022 7:40:14 AM | | CUDA: NVIDIA GPU 2 (not used): NVIDIA RTX A4500 (driver version 527.27, CUDA version 12.0, compute capability 8.6, 20470MB, 20470MB available, 23654 GFLOPS peak)
Why is it telling me that it is "not used"? It is showing up in the NVIDIA Control Panel and is listed as "use for graphic and compute".
I am not having this issue on any of the other workstations with the same driver.
by default, with a system containing mismatched GPUs, BOINC will only use the "best" GPU in the system. you can get around this with the use_all_gpus flag in your cc_config.xml
https://www.tomshardware.com/
)
https://www.tomshardware.com/news/sapphire-rx-6700-10gb-299-dollars">Rx6700 review. $300 Black Friday deal
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Not sure what an AMD card is
)
Not sure what an AMD card is doing in the “All things Nvidia” thread.
and your link on the face looks to be for some tomshardware article, but it just links back to your own post? What did you do there?
_________________________________________________________________________
Few days ago I had the chance
)
Few days ago I had the chance to try NVIDIA RTX 4090 Founders Edition in Einstein@Home and oh boy it literally chewed up those Gamma ray pulsar Work Units in a minute. I ran it for 1 hour and 30 minutes and Einstein@Home granted me 155K points so far and it keeps increasing every day.
Interestingly it only ran Gamma ray work units and nothing else. And work units seemed frozen at 87% and then get completed...
Nautilus wrote:... work units
)
For why this happens, you could try reading here.
The clue is in the thread title :-).
I think it's all about to change. The new 4000L series of tasks may have a much shorter follow-up stage, based on past behaviour. Overall, they may take a little longer to crunch since the 0-~90% stage for 4000L series used to take measurably longer.
Cheers,
Gary.
Nautilus wrote:work units
)
There's basically two parts to a work unit. Firstly there is the bulk of the work in producing a ( fast ) Fourier Transform from a given time series of measurements. Secondly the ( candidate ) signals - one of which may be a discovery - from the first stage must be ranked and sorted into a list to be returned to the project. So the % measure is certainly not a time indicator, as you have discovered. In fact I don't know/remember which part of the code does emit the % amount, nor how it calculates that. Does anyone know ?
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
I also noticed only half of
)
I also noticed only half of the GPU was being utilized and it was fairly cool (50ish celcius degrees)
Nautilus wrote: I also
)
That means you could try running more than a single task at a time and see if two task running at the same time take less time that each task running individually, if so and you still have headroom left you can try 3 or even more tasks at a time. Remember though that each gpu task needs cpu time and just leaving a single cpu core free to do that may not be enough to run at full speed.
Hi Mike! You explained it
)
Hi Mike!
You explained it so right!
The a) set-up, b) FFT, c) find signal candidates (abc-abc...) loop is done continuously in most Einstein task types and on some other projects (setiathome) too. i.e. the MAIN-loop.
Sometimes the after processing is done on CPU after the MAIN-loop and then you see a halt in progress percentage. That slow-down could be faked to seem like progress, but it actually represents the time it takes to do a similar amount of (flops) work.
With Nvidia and Cuda Linux executables you may be able to use cuda-mps-server to run multiple program executables in a (super) parallel manner to overcome the pre- and post processing halt. See the NVIDIA documentation.
--
petri
Hello all, Question- I
)
Hello all,
Question- I just updated the NVIDIA drivers and I received an odd message in BOINC that I am not sure I understand. I am not sure if it is related to the new drivers or the BOINC build I also recently updated.
11/29/2022 7:40:14 AM | | CUDA: NVIDIA GPU 2 (not used): NVIDIA RTX A4500 (driver version 527.27, CUDA version 12.0, compute capability 8.6, 20470MB, 20470MB available, 23654 GFLOPS peak)
Why is it telling me that it is "not used"? It is showing up in the NVIDIA Control Panel and is listed as "use for graphic and compute".
I am not having this issue on any of the other workstations with the same driver.
by default, with a system
)
by default, with a system containing mismatched GPUs, BOINC will only use the "best" GPU in the system. you can get around this with the use_all_gpus flag in your cc_config.xml
add <use_all_gpus>1</use_all_gpus>
to the options section of your cc_config.xml
_________________________________________________________________________