Yes you're right boinc is my main focus. I only have space for 1 GPU. There is a key difference between my computer and Ian&Steve C. is I am running Windows & Ian and Steve are running Linux. I believe Linux will always run work faster (thanks to the special app) than Windows for Einstein. I am not wanting to start a OS debate.
Well that's a tough one, among the top 50 list (https://einsteinathome.org/community/stats/hosts) where you can see OS and GPU used in the list I found 2 systems that use the AMD and NVIDIA card, both on Linux tough. On the plus side both use the same CPU so no CPU bias. Didn't see any of those cards on a Windows system. But on Linux the AMD Card is a bit faster.
Here are the computers, if you check their tasks you can verify for yourself. Of course you can't see if one is running them at 1x and the other at 2x but since the cards are in the same computing range and the computing time is similar I assume both use the same settings:
Of course what you don't see here is the power draw, I don't know which card is more efficient. And of course we don't know how this translates to Windows. If I had to decide I'd go for the card with bigger memory bandwidth as traditionally this is important in this project in the long run.
Can I interpret is as that higher CPU single-core performance matters a lot?
It depends.
Your initial question was:
Quote:
I'm considering buying a new computer. Is EAH FP64-intense? If so NVIDIA cards are not a good choice. More precisely, is 4060 better than 6700XT?
You should specify (in detail) what searches you intend to run as they tend to have quite different needs. Since you only mention a GPU comparison, are you intending to run GPU searches only or do you intend to have the available CPU threads running CPU tasks at the same time? My assumption is that you are mainly interested in GPU performance.
At the moment the two GPU searches are BRP7 and O3AS.
BRP7 uses uses very little CPU support and there is little difference in run time even if that support comes from an old/slow CPU. However, a GPU with better DP performance seems to make a reasonable reduction in the overall run time.
O3AS needs significant CPU involvement. A modern/fast CPU will make a significant difference to the run time. I have no information (yet) about the effect of a GPU with better DP performance but I expect it may not have much effect. There will be a significant benefit (whichever GPU type you choose) if you run concurrent tasks and stagger the start times of each so that (as much as possible) the GPU is always running one of the tasks whilst the other is in a 'heavy CPU use' mode. If both are running on the CPU at the same time (and the GPU is idle as a result) you waste resources and have longer run times.
Yes, I mean O3AS. I have tried to stagger the start time of the tasks but here'are still some problems.
First, BOINC manager switches between projects from time to time, and after switching to some project else and switching back, the staggering fails.
Second, my 4060 has only 8GB VRAM and can only support 2 concurrent tasks. That's not nice enough.
Third, even on newest CPUs, to reduce the CPU-only period, one need to run his CPU at very high frequencies. The sweet point of frequency of Raptor Lake is about 3GHz, which is still not enough for O3AS. Can it run multi-core? Or can it just tell BOINC manager to switch to CPU-only and free the VRAM use?
My initial guess is that, on 4060, the card run at full GPU load and maximum frequency but low power(in comparison to IOPs-intense project like PrimeGrid and local AI training). I turn to suppose that it is caused by that 4060 has little FP64 ability. Am I right?
I would most likely will run whatever works faster on the card I get. In this particular case probably BRP7 as that is where the focus is turning to, I believe.
I will most likely be running other things on my Ryzen 9 7900X CPU, I will always leave one core free to feed the GPU.
If you are running tasks on your CPU & have nothing running on your GPU how can this lead to longer run times? You may have a completely different project running on your GPU for example Amicable Numbers
In that case maybe I should rephrase my question and ask which card has better DP performance?
... BOINC manager switches between projects from time to time
Do you run separate projects each with a GPU search, or are you referring to separate searches at the one project?
On the assumption that you do indeed run GPU searches at several different projects, then you are going to lose production at the times when BOINC decides to switch between different projects. If you want to maximise the efficient use of a single computer for different projects with GPU searches, the simplest way would probably be to set all but one to NNT (No New Tasks) and rotate between projects at suitable intervals, eg. weekly or monthly, whatever is suitable, clearing the task list each time. That way you have a much better chance for optimising hardware use for each separate project without interference due to rapid unscheduled switching.
I tried to look at your hardware details but your computers are 'hidden'. Either change your prefs to allow non-sensitive hardware details to be seen or spell out full hardware specs for your system. This would allow better informed answers.
If you run multiple projects, then give a full list of them and what type of searches you run at each one. I only run Einstein but there are others who would be much better informed about the apps that run at other projects.
Or go with another Linux version, like Linux Mint, there's no unattended updates that I'm aware of.
That will give me something to consider. Thanks.
No need to switch distro - you can disable/uninstall unattended-upgrades
But as security updates are important, I would leave unattended-upgrades enabled and just pin the current driver version as per Ian&Steve C.'s comment and then do the driver updates manually (shut down BOINC, update driver, reboot).
Nvidia drivers should still be updated regularly as they contain security fixes, e.g. see here.
Regarding GPU choice, if I were to buy a new GPU I would always go with an AMD card as long as it works with Einstein@home, even if it were less efficient or harder to setup for crunching.
Good open source drivers available
AMD technologies are usually open, while Nvidia stuff is proprietary, e.g.
FSR vs DLSS
Freesync vs G-Sync
Cuda is proprietary and causes vendor lock-in, and I don't want to support that
I would most likely will run whatever works faster on the card I get. In this particular case probably BRP7 as that is where the focus is turning to, I believe.
The primary focus of the project has always been on the detection of continuous GW from spinning massive objects like neutron stars and black holes. It would be a really big deal if such a detection could be made. A non-detection is also important since it helps refine the parameters for such a phenomenon to occur.
Because LIGO goes through observation periods followed by non-observing periods where upgrades to improve sensitivity are made, the secondary searches like FGRP and various flavours of BRP serve the purpose of keeping the volunteers engaged when no LIGO observations are being performed. That's not meant to diminish the importance of secondary searches. However when recent observation data like O3 is available, I'm sure the project staff would want as much of that as possible to be done quickly.
This is basically why the run parameters have been modified for the v1.07 app which now doesn't exclude GPUs with 4GB VRAM. This allows a whole lot more machines to contribute that were previously unable to.
... BOINC manager switches between projects from time to time
Do you run separate projects each with a GPU search, or are you referring to separate searches at the one project?
On the assumption that you do indeed run GPU searches at several different projects, then you are going to lose production at the times when BOINC decides to switch between different projects. If you want to maximise the efficient use of a single computer for different projects with GPU searches, the simplest way would probably be to set all but one to NNT (No New Tasks) and rotate between projects at suitable intervals, eg. weekly or monthly, whatever is suitable, clearing the task list each time. That way you have a much better chance for optimising hardware use for each separate project without interference due to rapid unscheduled switching.
I tried to look at your hardware details but your computers are 'hidden'. Either change your prefs to allow non-sensitive hardware details to be seen or spell out full hardware specs for your system. This would allow better informed answers.
If you run multiple projects, then give a full list of them and what type of searches you run at each one. I only run Einstein but there are others who would be much better informed about the apps that run at other projects.
Nice suggestion! What I am doing now is just the opposite: I simply stopped O3AS :)
I will always leave one core free to feed the GPU.
By my experience GPU performance benefits from 1 core dedicated to each GPU task + 1 core completely free. So that one free core can handle system load, even if the machine is only used for crunching and nothing else I noticed that if I utilise all cores GPU run times prolong by a few minutes.
Speedy wrote: Yes you're
)
Well that's a tough one, among the top 50 list (https://einsteinathome.org/community/stats/hosts) where you can see OS and GPU used in the list I found 2 systems that use the AMD and NVIDIA card, both on Linux tough. On the plus side both use the same CPU so no CPU bias. Didn't see any of those cards on a Windows system. But on Linux the AMD Card is a bit faster.
Here are the computers, if you check their tasks you can verify for yourself. Of course you can't see if one is running them at 1x and the other at 2x but since the cards are in the same computing range and the computing time is similar I assume both use the same settings:
RX7900 XTX
https://einsteinathome.org/host/13105647
RTX 4080
https://einsteinathome.org/host/13071171
Of course what you don't see here is the power draw, I don't know which card is more efficient. And of course we don't know how this translates to Windows. If I had to decide I'd go for the card with bigger memory bandwidth as traditionally this is important in this project in the long run.
Can I interpret is as that
)
Can I interpret is as that higher CPU single-core performance matters a lot? If so should I choose non-X3D CPUs or X3D CPUs?
seewo wrote:Can I interpret
)
It depends.
Your initial question was:
You should specify (in detail) what searches you intend to run as they tend to have quite different needs. Since you only mention a GPU comparison, are you intending to run GPU searches only or do you intend to have the available CPU threads running CPU tasks at the same time? My assumption is that you are mainly interested in GPU performance.
At the moment the two GPU searches are BRP7 and O3AS.
BRP7 uses uses very little CPU support and there is little difference in run time even if that support comes from an old/slow CPU. However, a GPU with better DP performance seems to make a reasonable reduction in the overall run time.
O3AS needs significant CPU involvement. A modern/fast CPU will make a significant difference to the run time. I have no information (yet) about the effect of a GPU with better DP performance but I expect it may not have much effect. There will be a significant benefit (whichever GPU type you choose) if you run concurrent tasks and stagger the start times of each so that (as much as possible) the GPU is always running one of the tasks whilst the other is in a 'heavy CPU use' mode. If both are running on the CPU at the same time (and the GPU is idle as a result) you waste resources and have longer run times.
Cheers,
Gary.
Yes, I mean O3AS. I have
)
Yes, I mean O3AS. I have tried to stagger the start time of the tasks but here'are still some problems.
First, BOINC manager switches between projects from time to time, and after switching to some project else and switching back, the staggering fails.
Second, my 4060 has only 8GB VRAM and can only support 2 concurrent tasks. That's not nice enough.
Third, even on newest CPUs, to reduce the CPU-only period, one need to run his CPU at very high frequencies. The sweet point of frequency of Raptor Lake is about 3GHz, which is still not enough for O3AS. Can it run multi-core? Or can it just tell BOINC manager to switch to CPU-only and free the VRAM use?
My initial guess is that, on 4060, the card run at full GPU load and maximum frequency but low power(in comparison to IOPs-intense project like PrimeGrid and local AI training). I turn to suppose that it is caused by that 4060 has little FP64 ability. Am I right?
I would most likely will run
)
I would most likely will run whatever works faster on the card I get. In this particular case probably BRP7 as that is where the focus is turning to, I believe.
I will most likely be running other things on my Ryzen 9 7900X CPU, I will always leave one core free to feed the GPU.
If you are running tasks on your CPU & have nothing running on your GPU how can this lead to longer run times? You may have a completely different project running on your GPU for example Amicable Numbers
In that case maybe I should rephrase my question and ask which card has better DP performance?
seewo wrote:... BOINC manager
)
Do you run separate projects each with a GPU search, or are you referring to separate searches at the one project?
On the assumption that you do indeed run GPU searches at several different projects, then you are going to lose production at the times when BOINC decides to switch between different projects. If you want to maximise the efficient use of a single computer for different projects with GPU searches, the simplest way would probably be to set all but one to NNT (No New Tasks) and rotate between projects at suitable intervals, eg. weekly or monthly, whatever is suitable, clearing the task list each time. That way you have a much better chance for optimising hardware use for each separate project without interference due to rapid unscheduled switching.
I tried to look at your hardware details but your computers are 'hidden'. Either change your prefs to allow non-sensitive hardware details to be seen or spell out full hardware specs for your system. This would allow better informed answers.
If you run multiple projects, then give a full list of them and what type of searches you run at each one. I only run Einstein but there are others who would be much better informed about the apps that run at other projects.
Cheers,
Gary.
cecht wrote: JohnDK
)
No need to switch distro - you can disable/uninstall unattended-upgrades
But as security updates are important, I would leave unattended-upgrades enabled and just pin the current driver version as per Ian&Steve C.'s comment and then do the driver updates manually (shut down BOINC, update driver, reboot).
Nvidia drivers should still be updated regularly as they contain security fixes, e.g. see here.
Regarding GPU choice, if I were to buy a new GPU I would always go with an AMD card as long as it works with Einstein@home, even if it were less efficient or harder to setup for crunching.
If the card is also used for gaming, I would at a minimum get the 6700XT for the 12GB of VRAM (8GB is not enough anymore in modern games).
Speedy wrote:I would most
)
The primary focus of the project has always been on the detection of continuous GW from spinning massive objects like neutron stars and black holes. It would be a really big deal if such a detection could be made. A non-detection is also important since it helps refine the parameters for such a phenomenon to occur.
Because LIGO goes through observation periods followed by non-observing periods where upgrades to improve sensitivity are made, the secondary searches like FGRP and various flavours of BRP serve the purpose of keeping the volunteers engaged when no LIGO observations are being performed. That's not meant to diminish the importance of secondary searches. However when recent observation data like O3 is available, I'm sure the project staff would want as much of that as possible to be done quickly.
This is basically why the run parameters have been modified for the v1.07 app which now doesn't exclude GPUs with 4GB VRAM. This allows a whole lot more machines to contribute that were previously unable to.
Cheers,
Gary.
Gary Roberts wrote: seewo
)
Nice suggestion! What I am doing now is just the opposite: I simply stopped O3AS :)
Speedy wrote: I will always
)
By my experience GPU performance benefits from 1 core dedicated to each GPU task + 1 core completely free. So that one free core can handle system load, even if the machine is only used for crunching and nothing else I noticed that if I utilise all cores GPU run times prolong by a few minutes.