I have an Nvidia 780 and it is MUCH faster then my AMD 7970 that was crunching here!!
Quote:
I need to correct this, I do NOT have a 780, I have a 760 gpu!!
Sorry to disagree Mikey but, there must have been something very wrong with your configuration when running your 7970 here. Check out the Top Host lists.
7970/R9-280X cards running the openCL app versions perform far better than the equivalent NVidia Cuda app running cards (at least, here @Einstein) :-)
I have an Nvidia 780 and it is MUCH faster then my AMD 7970 that was crunching here!!
Quote:
I need to correct this, I do NOT have a 780, I have a 760 gpu!!
Sorry to disagree Mikey but, there must have been something very wrong with your configuration when running your 7970 here. Check out the Top Host lists.
7970/R9-280X cards running the openCL app versions perform far better than the equivalent NVidia Cuda app running cards (at least, here @Einstein) :-)
Gavin.
That doesn't really surprise me I guess, I did a quick and dirty attach and crunch and then didn't do much follow up as to how long each unit was taking as I have had to stop crunching here already, at least for now. I AM still crunching just not here right now, being on a Team means responsibilities and NOT passing an also crunching teammate just because you can isn't always a good thing!
As for your old Nvidia 660 doing better then your new R9-270 that is expected at Einstein, Einstein favors Nvidia gpu's in their programming. I have an Nvidia 780 and it is MUCH faster then my AMD 7970 that was crunching here!!
I need to correct this, I do NOT have a 780, I have a 760 gpu!!
Mikey, there is a bit more that needs correcting... 'Einstein favors NV' just doesn't stand up against the stats. The following is a Top Host stat summary from mid day yesterday:
The stats have been split by single card and multi-card hosts. Single card hosts tell the real story - AMD dominate at E@H, in spite of being outnumbered 2.5:1 by NV hosts! The highest ranked single card NV hosts are a 670 @ 91 overall ranking, a 570 @ 93 and a 670 @ 100. The highest ranked single card Titan/7XX are a 780Ti @ 105, a 770 @ 109 and a 780 @ 117.
As for your old Nvidia 660 doing better then your new R9-270 that is expected at Einstein, Einstein favors Nvidia gpu's in their programming. I have an Nvidia 780 and it is MUCH faster then my AMD 7970 that was crunching here!!
I need to correct this, I do NOT have a 780, I have a 760 gpu!!
Mikey, there is a bit more that needs correcting... 'Einstein favors NV' just doesn't stand up against the stats. The following is a Top Host stat summary from mid day yesterday:
The stats have been split by single card and multi-card hosts. Single card hosts tell the real story - AMD dominate at E@H, in spite of being outnumbered 2.5:1 by NV hosts! The highest ranked single card NV hosts are a 670 @ 91 overall ranking, a 570 @ 93 and a 670 @ 100. The highest ranked single card Titan/7XX are a 780Ti @ 105, a 770 @ 109 and a 780 @ 117.
Top 20 overall ranking favors AMD @ 75%...
Gord
I'd like to notice that you can't just take ranks and make some conclusion. I suppose we can compare only computers that are working 24/7. For example, please compare following positions:
94 with RAC 63,836.11 and NVIDIA GeForce GTX 570 (1279MB)
97 with RAC 62,745.21 and AMD Radeon HD 7870/7950/7970 series (Tahiti)
It would be wrong to say first one is faster then second because we don't know are those computers working full day or not.
94 with RAC 63,836.11 and NVIDIA GeForce GTX 570 (1279MB)
97 with RAC 62,745.21 and AMD Radeon HD 7870/7950/7970 series (Tahiti)
It would be wrong to say first one is faster then second because we don't know are those computers working full day or not.
Hi Sid,
You can't tell whether hosts are 24/7 directly. Digging through a host's task list and summing up the numbers is tedious and even so, doesn't tell the whole story. There are other factors that some into play.
An example, ranking 94 from your msg has current results for E@H only, whereas ranking 97 is active in close to 20 projects and we have no way of knowing what else is running on this host. What we do know is that a 7970 is capable of 2x the indicated RAC under the right conditions, so this host is not 100% E@H.
To get truly meaningful performance numbers for comparison you do have to search out hosts that are single project and running only one type of task. Any thing else produces muddy results...
Probably, we don't need to dig so comprehensively. Your computer list can give enough if you will be so kind to provide some information if they working in comparable conditions?
It would be wrong to say first one is faster then second because we don't know are those computers working full day or not.
I think the intent was to average over this "noise" (don't know Einstein ressource share, don't know clock speeds etc.) by considering the top 100 hosts. If there is any host running a card 24/7 for Einstein it should show up here - or the card is either slow or under-represented (like e.g. Titan Black).
Wuprop won't help here either because it only logs runtimes, but you can't tell how many WUs someone has been running in parallel.
Sid, which PCIe slot is your card placed in? (hosts are hidden)
I think the intent was to average over this "noise" (don't know Einstein ressource share, don't know clock speeds etc.) by considering the top 100 hosts.
Well, it's even worse than that.
The machine in the list with 4 NVIDIA 560Ti cards is mine and doesn't even have 4 560Tis in it. It's really two 470s and a 560Ti 448 and a "normal" SOC 560Ti. All on PCIe 2.0 running at x8 on a motherboard that can do that.
No CPU tasks are running.
It has been running Einstein 24/7 for a while. edit: BRP4 almost exclusively, for a while.
MY take, and I'm not prepared to do the research to give good numbers, is that my AMD cards are better at Einstein tasks (particularly Perseus Arm) on a per dollar and per watt basis.
BUT --- you can't really say it is "the card." The NVIDIA cards all want CUDA 4.2 and run really poorly on CUDA 3.2 on other projects. There would be a *significant* speed increase for all 400, 500, and 600 cards (I can't say about 700s) if they were running the "right" CUDA for them to be optimized.
My 660Tis and 670s get a smaller, but still significant speed increase from apps in CUDA 5 over CUDA 4.2. (at other projects with those applications)
So... I don't think there's any question that AMD cards run the *current* applications (Open CL) better than the NVIDIAs run the CUDA 3.2 which should have been re-written when the 400-series cards were released.
I understand there is an incompatibility problem between Einstein's code and some CUDA 4.2's something.
IF a CUDA 5 or 5.5 application were written, the leadership might well change and the project might find a 20%-30% increase in overall effective "FLOPS."
Anyway, just to make it CLEAR - I believe the AMD cards are superior here, running the code available here, available today.
Edit: I hope everyone knows a single GTX 690 is a dual GPU card and shows-up in the list as 2 cards; hence 4 GTX 690s are really 2 GTX 690s.
Sid, which PCIe slot is your card placed in? (hosts are hidden)
MrS
My card is placed to the PCI 2.0 16 slot. As Far as I know PCI 3.0 could speed up a bit -up to dozen of percents for BRP5 tasks.
Finally(I hope) I took MSI 770 card - it is a bit faster then my previous 560 Ti and all tasks have been validated so far.
Probably CUDA 5.X will solve the performance issue someday.
Sure it would be nice if there was an update of the CUDA app. But don't overestimate its benefit - it really depends on the code being used. CUDA is "just" the API, whereas the optimization for the newer architectures is performed by the GPU driver itself (otherwise old CUDA apps wouldn't work at all with new chips).
A performance improvement due to newer CUDA version can happen in two cases:
- New functionality is introduced, which helps to achieve some task more efficiently. You have to adapt your program to use this.
- Built-in functions are optimized. This works automatically nad is probably where the gains come from which you saw elsewhere. It's also what sometimes breaks things at GPU-Grid. But the benefit ultimately depends on you actually using CUDA built-in functions heavily. If I remember correctly Einstein spends most crunching time in a customized FFT, which would probably not automatically benefit from newer CUDA versions.
However, lacking any better hard numbers your guess is as good as mine ;)
RE: I have an Nvidia 780
)
Sorry to disagree Mikey but, there must have been something very wrong with your configuration when running your 7970 here. Check out the Top Host lists.
7970/R9-280X cards running the openCL app versions perform far better than the equivalent NVidia Cuda app running cards (at least, here @Einstein) :-)
Gavin.
RE: RE: I have an Nvidia
)
That doesn't really surprise me I guess, I did a quick and dirty attach and crunch and then didn't do much follow up as to how long each unit was taking as I have had to stop crunching here already, at least for now. I AM still crunching just not here right now, being on a Team means responsibilities and NOT passing an also crunching teammate just because you can isn't always a good thing!
RE: As for your old Nvidia
)
Mikey, there is a bit more that needs correcting... 'Einstein favors NV' just doesn't stand up against the stats. The following is a Top Host stat summary from mid day yesterday:
The stats have been split by single card and multi-card hosts. Single card hosts tell the real story - AMD dominate at E@H, in spite of being outnumbered 2.5:1 by NV hosts! The highest ranked single card NV hosts are a 670 @ 91 overall ranking, a 570 @ 93 and a 670 @ 100. The highest ranked single card Titan/7XX are a 780Ti @ 105, a 770 @ 109 and a 780 @ 117.
Top 20 overall ranking favors AMD @ 75%...
Gord
RE: RE: As for your old
)
I'd like to notice that you can't just take ranks and make some conclusion. I suppose we can compare only computers that are working 24/7. For example, please compare following positions:
94 with RAC 63,836.11 and NVIDIA GeForce GTX 570 (1279MB)
97 with RAC 62,745.21 and AMD Radeon HD 7870/7950/7970 series (Tahiti)
It would be wrong to say first one is faster then second because we don't know are those computers working full day or not.
However, I agree that AMD cards are in the top.
RE: 94 with RAC 63,836.11
)
Hi Sid,
You can't tell whether hosts are 24/7 directly. Digging through a host's task list and summing up the numbers is tedious and even so, doesn't tell the whole story. There are other factors that some into play.
An example, ranking 94 from your msg has current results for E@H only, whereas ranking 97 is active in close to 20 projects and we have no way of knowing what else is running on this host. What we do know is that a 7970 is capable of 2x the indicated RAC under the right conditions, so this host is not 100% E@H.
To get truly meaningful performance numbers for comparison you do have to search out hosts that are single project and running only one type of task. Any thing else produces muddy results...
Gord
RE: Any thing else
)
Probably, we don't need to dig so comprehensively. Your computer list can give enough if you will be so kind to provide some information if they working in comparable conditions?
RE: It would be wrong to
)
I think the intent was to average over this "noise" (don't know Einstein ressource share, don't know clock speeds etc.) by considering the top 100 hosts. If there is any host running a card 24/7 for Einstein it should show up here - or the card is either slow or under-represented (like e.g. Titan Black).
Wuprop won't help here either because it only logs runtimes, but you can't tell how many WUs someone has been running in parallel.
Sid, which PCIe slot is your card placed in? (hosts are hidden)
MrS
Scanning for our furry friends since Jan 2002
RE: I think the intent was
)
Well, it's even worse than that.
The machine in the list with 4 NVIDIA 560Ti cards is mine and doesn't even have 4 560Tis in it. It's really two 470s and a 560Ti 448 and a "normal" SOC 560Ti. All on PCIe 2.0 running at x8 on a motherboard that can do that.
No CPU tasks are running.
It has been running Einstein 24/7 for a while. edit: BRP4 almost exclusively, for a while.
MY take, and I'm not prepared to do the research to give good numbers, is that my AMD cards are better at Einstein tasks (particularly Perseus Arm) on a per dollar and per watt basis.
BUT --- you can't really say it is "the card." The NVIDIA cards all want CUDA 4.2 and run really poorly on CUDA 3.2 on other projects. There would be a *significant* speed increase for all 400, 500, and 600 cards (I can't say about 700s) if they were running the "right" CUDA for them to be optimized.
My 660Tis and 670s get a smaller, but still significant speed increase from apps in CUDA 5 over CUDA 4.2. (at other projects with those applications)
So... I don't think there's any question that AMD cards run the *current* applications (Open CL) better than the NVIDIAs run the CUDA 3.2 which should have been re-written when the 400-series cards were released.
I understand there is an incompatibility problem between Einstein's code and some CUDA 4.2's something.
IF a CUDA 5 or 5.5 application were written, the leadership might well change and the project might find a 20%-30% increase in overall effective "FLOPS."
Anyway, just to make it CLEAR - I believe the AMD cards are superior here, running the code available here, available today.
Edit: I hope everyone knows a single GTX 690 is a dual GPU card and shows-up in the list as 2 cards; hence 4 GTX 690s are really 2 GTX 690s.
RE: Sid, which PCIe slot
)
My card is placed to the PCI 2.0 16 slot. As Far as I know PCI 3.0 could speed up a bit -up to dozen of percents for BRP5 tasks.
Finally(I hope) I took MSI 770 card - it is a bit faster then my previous 560 Ti and all tasks have been validated so far.
Probably CUDA 5.X will solve the performance issue someday.
Sure it would be nice if
)
Sure it would be nice if there was an update of the CUDA app. But don't overestimate its benefit - it really depends on the code being used. CUDA is "just" the API, whereas the optimization for the newer architectures is performed by the GPU driver itself (otherwise old CUDA apps wouldn't work at all with new chips).
A performance improvement due to newer CUDA version can happen in two cases:
- New functionality is introduced, which helps to achieve some task more efficiently. You have to adapt your program to use this.
- Built-in functions are optimized. This works automatically nad is probably where the gains come from which you saw elsewhere. It's also what sometimes breaks things at GPU-Grid. But the benefit ultimately depends on you actually using CUDA built-in functions heavily. If I remember correctly Einstein spends most crunching time in a customized FFT, which would probably not automatically benefit from newer CUDA versions.
However, lacking any better hard numbers your guess is as good as mine ;)
MrS
Scanning for our furry friends since Jan 2002