No BOINC project apps have been written yet to make use of the Tensor cores in the latest Nvidia generations.
Only the general compute pipelines are used.
Maybe in the future, if part of the science/algorithm search can cope with lower precision requirements.
A little bit how we did when we changed from CPU only search to GPU search.
Not all search will be suited for it, but may be some can benefit from the greater speed.
We could do a 1º pass at the data to find potential candidates with lower precision, ant then re-rerun only the potential candidate with full precission.
Is this not a little bit what is done right now by the FP64 calculations at the end of each work unit?
I don't know exactly how the Tensor cores work. I only know that they are for performing simple mixed-precision matrix array maths.
The app devs would need to figure out if any part of their calculations would be speeded up with the tensor cores.
apparently Ampere (and later) GPUs can use the Tensor cores automatically for FP32 matrix operations. The way the nvidia documentation explains it, sounds like there's little or no code change necessary, but I think you have to be using CUDA libraries. that's all pretty vague so I don't know what the exact requirements are. but they call it "TF32". any matrix operations should be automatically picked up and run on tensor cores, but non-matrix will use the normal FP32 i think.
all of the examples I've seen have been pretty simple examples or purely AI/ML applications, so I'm not sure how this will translate to Einstein/BOINC or if it works with the CUDA app(s) here, or if they dont - how to get them working through code changes or compile changes. if there aren't a lot of matrix operations happening, it probably wont have any benefit.
I am about to replace two AMD GPUs with a new Nvidia RTX 4060 in a Ubuntu 22.04 system running a 6.2 Linux kernel. The display is plugged into the integrated GPU, so I only need compute functions on the 4060. If I simply install the 4060 will Ubuntu Drivers automatically install what is needed or do I need to go to Software & Updated > Additional Drivers, select, say, the 535-server proprietary package, then restart?
From another E@H discussion thread the <sudo apt install nvidia-driver-535> route was recommended, but I want to avoid restarting into a black screen, which I've been through in past years with AMD drivers (and what seems also to be a common problem in online Nvidia discussions). So is the apt or Ubuntu Drivers approach best for my particular setup and ingrained fears?
Either way, before installing or restarting, because I'm using on-board graphics, do I need to edit the grub file with GRUB_CMDLINE_LINUX_DEFAULT="quiet splash nomodeset" or "nouveau.modeset=0"?
Ideas are not fixed, nor should they be; we live in model-dependent reality.
I am about to replace two AMD GPUs with a new Nvidia RTX 4060 in a Ubuntu 22.04 system running a 6.2 Linux kernel. The display is plugged into the integrated GPU, so I only need compute functions on the 4060. If I simply install the 4060 will Ubuntu Drivers automatically install what is needed or do I need to go to Software & Updated > Additional Drivers, select, say, the 535-server proprietary package, then restart?
From another E@H discussion thread the <sudo apt install nvidia-driver-535> route was recommended, but I want to avoid restarting into a black screen, which I've been through in past years with AMD drivers (and what seems also to be a common problem in online Nvidia discussions). So is the apt or Ubuntu Drivers approach best for my particular setup and ingrained fears?
Either way, before installing or restarting, because I'm using on-board graphics, do I need to edit the grub file with GRUB_CMDLINE_LINUX_DEFAULT="quiet splash nomodeset" or "nouveau.modeset=0"?
I use Linux Mint and when I load it on a new pc it asks me if I want to install "additional drivers" during the setup questions but then I have to go into the additional drivers route to get it up to the 525 or 535 lvel of the drivers. It doesn't actually install Nvidia drivers during the adidtional setup just Nvidia compatible ones. I did upgrade one pc thru the additional drivers path to the 535 driver and I got the black screen and had to restart the whole process over which is why I'm on the 525 version. Now I do NOT use the latest version of Linux Mint because Boinc won't run, it loads just fine though and even the '....start' or '....stop' commands don't seem to do anything. LM Admins know about the problem but have bigger problems to fix.
Once I get the drivers loaded things work great and Boinc crunches with the Nvidia gpu with no problems.
I have using the gui via the os update program. It offers a variety of versions below 535. Select the proprietary version. Do not try to use the 390 version. I believe 470 works.
I have a Ryzen 2400G but stopped fighting with trying to use the internal GPU when I had a discrete GPU installed. I have no advice.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Filipe wrote: Will it make
)
I doubt it. Einstein is using FP32 & FP64 now and it doesn't make sense (at least to me) if they reverse the trend and go down to FP16 or FP8.
That's my take on it, ...and I'm stickin' to it! ;>)
Proud member of the Old Farts Association
No BOINC project apps have
)
No BOINC project apps have been written yet to make use of the Tensor cores in the latest Nvidia generations.
Only the general compute pipelines are used.
Keith Myers wrote:No BOINC
)
Maybe in the future, if part of the science/algorithm search can cope with lower precision requirements.
A little bit how we did when we changed from CPU only search to GPU search.
Not all search will be suited for it, but may be some can benefit from the greater speed.
We could do a 1º pass at the data to find potential candidates with lower precision, ant then re-rerun only the potential candidate with full precission.
Is this not a little bit what is done right now by the FP64 calculations at the end of each work unit?
I don't know exactly how the
)
I don't know exactly how the Tensor cores work. I only know that they are for performing simple mixed-precision matrix array maths.
The app devs would need to figure out if any part of their calculations would be speeded up with the tensor cores.
Keith Myers wrote: I don't
)
apparently Ampere (and later) GPUs can use the Tensor cores automatically for FP32 matrix operations. The way the nvidia documentation explains it, sounds like there's little or no code change necessary, but I think you have to be using CUDA libraries. that's all pretty vague so I don't know what the exact requirements are. but they call it "TF32". any matrix operations should be automatically picked up and run on tensor cores, but non-matrix will use the normal FP32 i think.
https://blogs.nvidia.com/blog/tensorfloat-32-precision-format/
all of the examples I've seen have been pretty simple examples or purely AI/ML applications, so I'm not sure how this will translate to Einstein/BOINC or if it works with the CUDA app(s) here, or if they dont - how to get them working through code changes or compile changes. if there aren't a lot of matrix operations happening, it probably wont have any benefit.
_________________________________________________________________________
Ian&SteveC, Speculating
)
Ian&SteveC,
Speculating here but if it was that automatic games wouldn't have had to be rewritten to use it?
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Tom M
)
it has nothing to do with games.
_________________________________________________________________________
I am about to replace two AMD
)
I am about to replace two AMD GPUs with a new Nvidia RTX 4060 in a Ubuntu 22.04 system running a 6.2 Linux kernel. The display is plugged into the integrated GPU, so I only need compute functions on the 4060. If I simply install the 4060 will Ubuntu Drivers automatically install what is needed or do I need to go to Software & Updated > Additional Drivers, select, say, the 535-server proprietary package, then restart?
From another E@H discussion thread the <sudo apt install nvidia-driver-535> route was recommended, but I want to avoid restarting into a black screen, which I've been through in past years with AMD drivers (and what seems also to be a common problem in online Nvidia discussions). So is the apt or Ubuntu Drivers approach best for my particular setup and ingrained fears?
Either way, before installing or restarting, because I'm using on-board graphics, do I need to edit the grub file with GRUB_CMDLINE_LINUX_DEFAULT="quiet splash nomodeset" or "nouveau.modeset=0"?
Ideas are not fixed, nor should they be; we live in model-dependent reality.
cecht wrote: I am about to
)
I use Linux Mint and when I load it on a new pc it asks me if I want to install "additional drivers" during the setup questions but then I have to go into the additional drivers route to get it up to the 525 or 535 lvel of the drivers. It doesn't actually install Nvidia drivers during the adidtional setup just Nvidia compatible ones. I did upgrade one pc thru the additional drivers path to the 535 driver and I got the black screen and had to restart the whole process over which is why I'm on the 525 version. Now I do NOT use the latest version of Linux Mint because Boinc won't run, it loads just fine though and even the '....start' or '....stop' commands don't seem to do anything. LM Admins know about the problem but have bigger problems to fix.
Once I get the drivers loaded things work great and Boinc crunches with the Nvidia gpu with no problems.
I have using the gui via the
)
I have using the gui via the os update program. It offers a variety of versions below 535. Select the proprietary version. Do not try to use the 390 version. I believe 470 works.
I have a Ryzen 2400G but stopped fighting with trying to use the internal GPU when I had a discrete GPU installed. I have no advice.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)