I was also curious about why CUDA55 as opposed to more recent CUDA versions. Perhaps there is just nothing in CUDA 6+ that E@H can take advantage of?
I think that's it. And Einstein has historically supported a large user base with a huge variety of systems. Moving to a newer CUDA version would cut some people out who are limited to older drivers for some reason. One could combat this by deploying multiple app versions for different CUDA versions, but this creates more work during building & validation. So it's always a trade-off and the choice has been CUDA 5.5 as a "golden standard" this time. It doesn't exclude too many crunchers but provides a significant gain over CUDA 3.2.
MrS
Please inform yourself before posting wrong things...
Quoting myself:
Quote:
The CUDA 6.0 toolkit depreciates targeting G80 (sm_10) but still supports this architecture. CUDA 6.5 removed sm_10 support but still supports sm_11, sm_12, and sm_13 architectures (although deprecated).
I can't imagine there are still a lot of G8x GPUs out there but who knows...
Thus, CUDA 6.0 does not exclude anyone. Even CUDA 6.5 would probably lock out only very few GPUs with minimal impact (G80 has only 128 CUDA cores!). As CUDA 6.0 introduced Maxwell support, it might be more interesting than CUDA 5.5.
Thus, CUDA 6.0 does not exclude anyone. Even CUDA 6.5 would probably lock out only very few GPUs with minimal impact (G80 has only 128 CUDA cores!). As CUDA 6.0 introduced Maxwell support, it might be more interesting than CUDA 5.5.
edit: This was discussed a little bit earlier above in this thread
I guess support for OS such as XP - might also be a factor?
... Perhaps there is just nothing in CUDA 6+ that E@H can take advantage of?
and MrS replied:
Quote:
I think that's it ....
to which |MatMan| responded:
Quote:
Please inform yourself before posting wrong things...
Quoting myself:
Quote:
The CUDA 6.0 toolkit depreciates targeting G80 (sm_10) but still supports this architecture. CUDA 6.5 removed sm_10 support but still supports sm_11, sm_12, and sm_13 architectures (although deprecated).
I can't imagine there are still a lot of G8x GPUs out there but who knows...
Thus, CUDA 6.0 does not exclude anyone. Even CUDA 6.5 would probably lock out only very few GPUs with minimal impact (G80 has only 128 CUDA cores!). As CUDA 6.0 introduced Maxwell support, it might be more interesting than CUDA 5.5.
A couple of points:-
1. There's no need to be rude.
2. If you are going to quote (a part of) a discussion that occurred quite a while ago, you should provide a link so people can easily find the full context, AND you should also acknowledge (and link to) the full answer you were given at the time as to why the Devs had decided to go only to CUDA 5.5 at this stage. By not doing so you give the impression, either you don't believe the answer, or you have completely forgotten about it (in which case follow your own advice), or the answer is inconvenient for the current points you want to make.
It shouldn't be necessary for other readers to find and provide the link themselves. When you read that full answer, courtesy of AgentB's link, it validates MrS' comments. It shows why this project performs as well as it does with the minimal staff available to run it. The Devs make every effort to ensure that things will work correctly without causing unnecessary grief and that there will be a benefit to be had from the changes when implemented.
1. There's no need to be rude.
2. If you are going to quote (a part of) a discussion that occurred quite a while ago, you should provide a link so people can easily find the full context, AND you should also acknowledge (and link to) the full answer you were given at the time as to why the Devs had decided to go only to CUDA 5.5 at this stage. By not doing so you give the impression, either you don't believe the answer, or you have completely forgotten about it (in which case follow your own advice), or the answer is inconvenient for the current points you want to make.
Sorry for being a bit rude, this was not my intention.
I still remember that discussion. The Devs also stated, that they wanted to test all the different CUDA versions up to 7.x for possible benefits. My point was not to question why we use CUDA 5.5 right now (which I didn't). I'm glad about the gains from CUDA 5.5!
The forum here doesn't allow to set links to quotes. So yes, I was too lazy to post an extra link. Sorry about that.
Quote:
It shouldn't be necessary for other readers to find and provide the link themselves. When you read that full answer, courtesy of AgentB's link, it validates MrS' comments. It shows why this project performs as well as it does with the minimal staff available to run it. The Devs make every effort to ensure that things will work correctly without causing unnecessary grief and that there will be a benefit to be had from the changes when implemented.
I was basically only referring to the very last part of MrS' statement:
Quote:
So it's always a trade-off and the choice has been CUDA 5.5 as a "golden standard" this time. It doesn't exclude too many crunchers but provides a significant gain over CUDA 3.2.
which is simply not true (see CUDA release notes, Deprecated Features on page 10). CUDA 5.5 or CUDA 6.0 do not exclude any CUDA-capable GPU.
So it's always a trade-off and the choice has been CUDA 5.5 as a "golden standard" this time. It doesn't exclude too many crunchers but provides a significant gain over CUDA 3.2.
which is simply not true (see CUDA release notes, Deprecated Features on page 10). CUDA 5.5 or CUDA 6.0 do not exclude any CUDA-capable GPU.
You seem to be a focussed on the GPU card being the only issue here.
It is not.
I will repeat my comments which you seem not to have read, about OS being unsupported by CUDA 6.0 as "possibly" being one of the issues.
If we assume the Nvidia release notes, as being the only source of information (these may or may not be complete or true), may i refer you (other readers may skip this) to CUDA 5.5 release notes
It categorically states there (i couldn't decide on red or green - i settled for Ubuntu red, apologies nVidia).
Ubuntu 10.04 LTS
We recommend upgrading to the Ubuntu 12.04 LTS. Support for this operating
system will be removed in the next release of the CUDA software.
... so that would be 6.0.
If you think, that matters not....
My long serving host - one of many Ubuntus out there in the wild- muon runs Ubuntu 10.04 LTS and has RAC 100K and a total over 55M.
So I would like to thank the devs for making a wise choice.
Thus, CUDA 6.0 does not exclude anyone. Even CUDA 6.5 would probably lock out only very few GPUs with minimal impact (G80 has only 128 CUDA cores!). As CUDA 6.0 introduced Maxwell support, it might be more interesting than CUDA 5.5.
edit: This was discussed a little bit earlier above in this thread
I guess support for OS such as XP - might also be a factor?
I'm running CUDA 6.5 under Windows XP. GPUGrid supply the application, NVidia supply the driver (344.75) and runtime support, and I supplied the GTX 750 Ti Maxwell card.
GPUGrid looked at CUDA 6.5 and decided that it had features beneficial to their applications. Einstein - I'm sure - have looked at it, and decided that it hasn't, for the time being at least. There's a bit of a difference between biochemistry and astrophysics, after all.
I have read your posts and I agree to disagree ;) I see the focus of excluding possible crunchers mostly on the hardware side. Software can be updated, often for free. Sure, that is my personal opinion.
Ubuntu LTS can also be upgraded. Recently, I've upgraded XUbuntu LTS to the latest version with minimal effort on a few of my hosts (basically just a mouse click). Ubuntu 10.04 LTS is more than 5 years old after all.
They way I understood the Devs, the plan was to keep a "compatibility-version" anyway (be it CUDA 3.2 or CUDA 5.5).
It categorically states there (i couldn't decide on red or green - i settled for Ubuntu red, apologies nVidia).
Ubuntu 10.04 LTS
We recommend upgrading to the Ubuntu 12.04 LTS. Support for this operating
system will be removed in the next release of the CUDA software.
... so that would be 6.0.
If you think, that matters not....
I was thinking about this issue again: it does not matter. It only matters for developers. You don't need to install the CUDA toolkit to run Einstein tasks. You only need a compatible driver. The driver of your Ubuntu 10.04 host supports even CUDA 7.0 (taken from a random task):
Looks like no one has commented on this thread in some time.
Reading through the post from July to August, there were several concerns over the v1.57 on Maxwell GPUs.
Keith made note that he thought the run times weren't any advantage over v1.52. If anything, he thought the times were longer.
But, reading his posts. He wasn't testing the work units by itself. He was using a mixed configuration of Seti and Einstein.
We already knew that the 2 different types of work units at Seti interfere with each other, so running them together was not beneficial to overall efficency.
So it comes as no surprise that doing the same with an Einstein work unit would cause it to slow down as well. Especially if you are running 2 Seti work units to 1 Einstein.
I like Keith (no disrespect to you) but to be fair to the test unit v1.57, it has to run by itself or in groups then compared to v1.52
So, on that note.
So far I've run over 3000 of them (closer to 4000) and have only 1 invalid. The 8 errors are due to a reinstall of GPUs and then M$ deciding to load it's own drivers on the cards (causing those work units in progress to error out)
There is significant improvement in v1.57 compared to v1.52
This increases with the number of work units running in parallel on a given GPU.
RE: RE: I was also
)
Please inform yourself before posting wrong things...
Quoting myself:
Thus, CUDA 6.0 does not exclude anyone. Even CUDA 6.5 would probably lock out only very few GPUs with minimal impact (G80 has only 128 CUDA cores!). As CUDA 6.0 introduced Maxwell support, it might be more interesting than CUDA 5.5.
RE: Thus, CUDA 6.0 does
)
edit: This was discussed a little bit earlier above in this thread
I guess support for OS such as XP - might also be a factor?
Matt said: RE: ...
)
Matt said:
and MrS replied:
to which |MatMan| responded:
A couple of points:-
1. There's no need to be rude.
2. If you are going to quote (a part of) a discussion that occurred quite a while ago, you should provide a link so people can easily find the full context, AND you should also acknowledge (and link to) the full answer you were given at the time as to why the Devs had decided to go only to CUDA 5.5 at this stage. By not doing so you give the impression, either you don't believe the answer, or you have completely forgotten about it (in which case follow your own advice), or the answer is inconvenient for the current points you want to make.
It shouldn't be necessary for other readers to find and provide the link themselves. When you read that full answer, courtesy of AgentB's link, it validates MrS' comments. It shows why this project performs as well as it does with the minimal staff available to run it. The Devs make every effort to ensure that things will work correctly without causing unnecessary grief and that there will be a benefit to be had from the changes when implemented.
Cheers,
Gary.
RE: A couple of
)
Sorry for being a bit rude, this was not my intention.
I still remember that discussion. The Devs also stated, that they wanted to test all the different CUDA versions up to 7.x for possible benefits. My point was not to question why we use CUDA 5.5 right now (which I didn't). I'm glad about the gains from CUDA 5.5!
The forum here doesn't allow to set links to quotes. So yes, I was too lazy to post an extra link. Sorry about that.
I was basically only referring to the very last part of MrS' statement:
which is simply not true (see CUDA release notes, Deprecated Features on page 10). CUDA 5.5 or CUDA 6.0 do not exclude any CUDA-capable GPU.
RE: RE: So it's always a
)
You seem to be a focussed on the GPU card being the only issue here.
It is not.
I will repeat my comments which you seem not to have read, about OS being unsupported by CUDA 6.0 as "possibly" being one of the issues.
If we assume the Nvidia release notes, as being the only source of information (these may or may not be complete or true), may i refer you (other readers may skip this) to CUDA 5.5 release notes
It categorically states there (i couldn't decide on red or green - i settled for Ubuntu red, apologies nVidia).
Ubuntu 10.04 LTS
We recommend upgrading to the Ubuntu 12.04 LTS. Support for this operating
system will be removed in the next release of the CUDA software.
... so that would be 6.0.
If you think, that matters not....
My long serving host - one of many Ubuntus out there in the wild- muon runs Ubuntu 10.04 LTS and has RAC 100K and a total over 55M.
So I would like to thank the devs for making a wise choice.
MrS is correct in his statement.
Postel's law applies.
RE: RE: Thus, CUDA 6.0
)
I'm running CUDA 6.5 under Windows XP. GPUGrid supply the application, NVidia supply the driver (344.75) and runtime support, and I supplied the GTX 750 Ti Maxwell card.
GPUGrid looked at CUDA 6.5 and decided that it had features beneficial to their applications. Einstein - I'm sure - have looked at it, and decided that it hasn't, for the time being at least. There's a bit of a difference between biochemistry and astrophysics, after all.
RE: You seem to be a
)
I have read your posts and I agree to disagree ;) I see the focus of excluding possible crunchers mostly on the hardware side. Software can be updated, often for free. Sure, that is my personal opinion.
Ubuntu LTS can also be upgraded. Recently, I've upgraded XUbuntu LTS to the latest version with minimal effort on a few of my hosts (basically just a mouse click). Ubuntu 10.04 LTS is more than 5 years old after all.
They way I understood the Devs, the plan was to keep a "compatibility-version" anyway (be it CUDA 3.2 or CUDA 5.5).
RE: It categorically states
)
I was thinking about this issue again: it does not matter. It only matters for developers. You don't need to install the CUDA toolkit to run Einstein tasks. You only need a compatible driver. The driver of your Ubuntu 10.04 host supports even CUDA 7.0 (taken from a random task):
Looks like no one has
)
Looks like no one has commented on this thread in some time.
Reading through the post from July to August, there were several concerns over the v1.57 on Maxwell GPUs.
Keith made note that he thought the run times weren't any advantage over v1.52. If anything, he thought the times were longer.
But, reading his posts. He wasn't testing the work units by itself. He was using a mixed configuration of Seti and Einstein.
We already knew that the 2 different types of work units at Seti interfere with each other, so running them together was not beneficial to overall efficency.
So it comes as no surprise that doing the same with an Einstein work unit would cause it to slow down as well. Especially if you are running 2 Seti work units to 1 Einstein.
I like Keith (no disrespect to you) but to be fair to the test unit v1.57, it has to run by itself or in groups then compared to v1.52
So, on that note.
So far I've run over 3000 of them (closer to 4000) and have only 1 invalid. The 8 errors are due to a reinstall of GPUs and then M$ deciding to load it's own drivers on the cards (causing those work units in progress to error out)
There is significant improvement in v1.57 compared to v1.52
This increases with the number of work units running in parallel on a given GPU.
Zalster