Advanced search

Message boards : Graphics cards (GPUs) : CUDA 7.5 drivers are faster than CUDA 8 on GTX 970

Author Message
Jim1348
Send message
Joined: 28 Jul 12
Posts: 462
Credit: 1,130,762,168
RAC: 13,533
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwat
Message 43750 - Posted: 8 Jun 2016 | 18:42:55 UTC
Last modified: 8 Jun 2016 | 18:46:04 UTC

A posting on another forum stated that some of the earlier Nvidia drivers are faster than later ones, though I don't have link to that post (probably on Einstein). So I tried it out on my GTX 970 on GPUGrid, going back from 361.43 (the first of the CUDA 8 drivers) to 359.06 (the last of the CUDA 7.5 drivers). I tried to find the most comparable work units, but the trend is clear whether you get an exact match or not (Win7 64-bit in all cases):

361.43 (CUDA 8) driver:
e7s29_e2s13p0f526-GERARD_CXCL12VOLK_12998741_2-0-1-RND3569_1
13.85 hours (44% CPU utilization)

359.06 (CUDA 7.5) driver:
e12s19_e7s8p0f14-GERARD_CXCL12VOLK_12998741_1-0-1-RND7922_0
12.59 hours (31% CPU utilization)

The conclusion is that the older drivers produce both faster results and lower CPU utilization. (This card is supported by a full core of an i5-3550, so the numbers will be somewhat different on virtual cores.)

While I was at it, I compared the efficiency (work units/day/watt) of the GTX 970 to a GTX 960. Somewhat to my surprise, they are virtually the same, with the GTX 960 squeaking out maybe a 2 or 3 percent gain, but almost too small to measure reliably.

Profile Beyond
Avatar
Send message
Joined: 23 Nov 08
Posts: 1073
Credit: 4,497,656,004
RAC: 407,908
Level
Arg
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 43751 - Posted: 8 Jun 2016 | 18:48:15 UTC - in response to Message 43750.
Last modified: 8 Jun 2016 | 18:49:35 UTC

Very interesting. What about the "CUDA 8" drivers that are newer, perhaps more mature?
How about even older drivers. Any ideas on those vs. 7.5?

Jim1348
Send message
Joined: 28 Jul 12
Posts: 462
Credit: 1,130,762,168
RAC: 13,533
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwat
Message 43752 - Posted: 8 Jun 2016 | 18:54:06 UTC - in response to Message 43751.

Good questions, but my experience is that the CUDA drivers are pretty stable once they are released, and the later driver versions just optimize gaming performance. I don't do games, and this is on a dedicated machine anyway, so that does not matter.

You are more than welcome to try it also, and post back.

By the way, I measured power using the TDP% in GPU-Z, and the CPU% in BoincTasks.

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2671
Credit: 751,855,924
RAC: 450,091
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 43766 - Posted: 11 Jun 2016 | 7:32:42 UTC

Same here:

362.00
41203 s for 280,650 credit WUs (6 samples)

359.06
38818 s for 280,650 credit WUs (3 samples)

MrS
____________
Scanning for our furry friends since Jan 2002

Profile skgiven
Volunteer moderator
Project tester
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,835,617,224
RAC: 310,037
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 43769 - Posted: 11 Jun 2016 | 18:25:09 UTC - in response to Message 43766.
Last modified: 21 Jun 2016 | 8:32:11 UTC

This has been seen before here.
Drivers become bloated with patches and it slows some apps down. Operating systems get bloated in a similar way. Basically each fix introduces an additional subroutine or set of subroutines.
The difference observed by Jim1348 was 10% while the difference observed by ETA was 6%, so perhaps subsequent drivers (ETA's being slightly newer) or different setups change things.

The latest apps hear are CUDA 6.5 so any drivers that support 6.5 should suffice (unless there has been reported issues).
To test any differences you do need a steady supply of the same work type, and to stick to the same setup.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Jim1348
Send message
Joined: 28 Jul 12
Posts: 462
Credit: 1,130,762,168
RAC: 13,533
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwat
Message 43780 - Posted: 14 Jun 2016 | 16:31:12 UTC - in response to Message 43769.

Drivers become bloated with patches and it slows some apps down. Operating systems get bloated in a similar way.

But the point here is that going from the last CUDA 7.5 driver to the first CUDA 8 shows that the difference is due mainly to the difference in the CUDA versions. That may be called "bloat", as there may be some additional overhead in CUDA 8 that is not needed, but otherwise your statement is rather general and not supported by any data or references to this effect.

The measurements that ETA got I expect are mainly due to the fact that he averaged over a larger number of work unit types. It would be nice to do only direct comparisons, but that may not be practical for a large number of WU's. But again, if someone wants to compare later (or earlier) drivers, that would be helpful too, as long as they state their methodology.

Profile Beyond
Avatar
Send message
Joined: 23 Nov 08
Posts: 1073
Credit: 4,497,656,004
RAC: 407,908
Level
Arg
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 43783 - Posted: 16 Jun 2016 | 2:10:50 UTC - in response to Message 43769.

Same here: 362.00
41203 s for 280,650 credit WUs (6 samples)

359.06
38818 s for 280,650 credit WUs (3 samples) MrS

The difference observed by Jim1348 was 10% while the difference observed by ETA was 2%, so perhaps subsequent drivers (ETA's being slightly newer) or different setups change things.

The latest apps hear are CUDA 6.5 so any drivers that support 6.5 should suffice (unless there has been reported issues).

Doesn't 41203/38818=1.06144 Or a bit over 6% ?



Profile skgiven
Volunteer moderator
Project tester
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,835,617,224
RAC: 310,037
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 43803 - Posted: 21 Jun 2016 | 9:14:33 UTC - in response to Message 43783.

Yes, thanks, should have said 362 took 6% longer & it does look like CUDA 8 drivers are slower at running CUDA 6.5 apps than CUDA 7.5 drivers.
Maybe I'm wrong but I presumed CUDA 8 was developed for Pascal and then patched for backward compatibility with earlier CUDA versions? Whatever the route I don't see why NV would dev for older products, so I would not expect any performance gain from CUDA 8 for 28nm GTX cards, though they might improve upon their earliest CUDA 8 drivers somewhat. As usual I suggest people stick to the CUDA 7.5 drivers they have unless they need new drivers or want to test and report on them - which is useful if it confirms CUDA 7.5 drivers are faster than CUDA 8 for GTX 900 series GPU's or any changes to the 6% to 10% performance drop.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Profile Beyond
Avatar
Send message
Joined: 23 Nov 08
Posts: 1073
Credit: 4,497,656,004
RAC: 407,908
Level
Arg
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 43806 - Posted: 21 Jun 2016 | 17:31:33 UTC - in response to Message 43803.

Whatever the route I don't see why NV would dev for older products, so I would not expect any performance gain from CUDA 8 for 28nm GTX cards

It would amount to customer loyalty, which unfortunately does not seem to be the current business model for NVidia and many other corporations.

Profile Beyond
Avatar
Send message
Joined: 23 Nov 08
Posts: 1073
Credit: 4,497,656,004
RAC: 407,908
Level
Arg
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 43807 - Posted: 21 Jun 2016 | 17:46:39 UTC

Jim, thanks for this heads up. Switching to the 359.06 drivers has allowed one of my machines to make the 24 hour deadline for 7 of the last 11 WUs (sometimes by as little as 2.5 minutes). Previously it was missing all the 50% bonus times after "upgrading" to 362 drivers. I'm currently not having warm fuzzy feelings about nvidia.

Erich56
Send message
Joined: 1 Jan 15
Posts: 372
Credit: 1,679,211,402
RAC: 2,917,345
Level
His
Scientific publications
watwatwat
Message 43823 - Posted: 23 Jun 2016 | 11:06:56 UTC - in response to Message 43807.

Switching to the 359.06 drivers has allowed one of my machines to make the 24 hour deadline for 7 of the last 11 WUs (sometimes by as little as 2.5 minutes). Previously it was missing all the 50% bonus times after "upgrading" to 362 drivers. I'm currently not having warm fuzzy feelings about nvidia.

could you please let us know which GPU and which OS you talk about in this case?

I am asking because I seem to be in a very similar situation with one my PCs.
Thanks in advance.

Jim1348
Send message
Joined: 28 Jul 12
Posts: 462
Credit: 1,130,762,168
RAC: 13,533
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwat
Message 43824 - Posted: 23 Jun 2016 | 11:18:03 UTC - in response to Message 43807.

Jim, thanks for this heads up. Switching to the 359.06 drivers has allowed one of my machines to make the 24 hour deadline for 7 of the last 11 WUs (sometimes by as little as 2.5 minutes). Previously it was missing all the 50% bonus times after "upgrading" to 362 drivers. I'm currently not having warm fuzzy feelings about nvidia.

You are welcome, and it surprised me a bit too. It was all due to someone on another forum though who noticed it first. I don't recall a comparable case for CUDA, though it used to happen all the time for the AMD OpenCl drivers.

Profile Beyond
Avatar
Send message
Joined: 23 Nov 08
Posts: 1073
Credit: 4,497,656,004
RAC: 407,908
Level
Arg
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 43825 - Posted: 23 Jun 2016 | 19:25:17 UTC - in response to Message 43823.
Last modified: 23 Jun 2016 | 19:26:11 UTC

Switching to the 359.06 drivers has allowed one of my machines to make the 24 hour deadline for 7 of the last 11 WUs (sometimes by as little as 2.5 minutes). Previously it was missing all the 50% bonus times after "upgrading" to 362 drivers. I'm currently not having warm fuzzy feelings about nvidia.

could you please let us know which GPU and which OS you talk about in this case?

I am asking because I seem to be in a very similar situation with one my PCs.
Thanks in advance.

Windows 7 and NV 750Ti factory super-clocked. It's only really working on 1 of my machines, the others are crunching just a bit slower so they don't quite make the 50% bonus. I suspect this is going to be short lived as WU sizes keep increasing even though GPUs have remained the same for quite a while. In my case upload time is also a factor as file sizes have increased. I've been looking at skgiven's Linux thread, a switch to Linux should yield about an 11% boost AFAIK.

Betting Slip
Send message
Joined: 5 Jan 09
Posts: 589
Credit: 2,041,743,975
RAC: 1,505,145
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 43826 - Posted: 23 Jun 2016 | 19:37:32 UTC - in response to Message 43825.

Maybe you should ALL think about contribution rather than speed and "beggar thy neighbour"

Jim1348
Send message
Joined: 28 Jul 12
Posts: 462
Credit: 1,130,762,168
RAC: 13,533
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwat
Message 43829 - Posted: 24 Jun 2016 | 13:02:34 UTC
Last modified: 24 Jun 2016 | 13:06:36 UTC

As luck would have it, my GTX 970 machine crashed yesterday for no good reason and would not boot at all. The SSD (a Samsung 850 EVO) even disappeared until I re-enabled it in the BIOS. But after completely wiping it out with a secure erase, it appeared to be OK. Rather than re-installing Win7 64-bit, I decided to go with Ubuntu 16.04; the second of my machines to be converted.

The first GPU work unit was a GERARD_CXCL12VOLKX. That may not be exactly comparable to the one I did on Windows with the CUDA 7.5 driver, but at 11.2 hours, a 12% difference, it shows that things are heading in the right direction. I would not bother with that except for the crash, but it was a nice incentive.

Erich56
Send message
Joined: 1 Jan 15
Posts: 372
Credit: 1,679,211,402
RAC: 2,917,345
Level
His
Scientific publications
watwatwat
Message 43830 - Posted: 24 Jun 2016 | 18:10:05 UTC - in response to Message 43826.

Maybe you should ALL think about contribution rather than speed and "beggar thy neighbour"

I know what you want to say - but on this particular PC, crunching time is some 24,5 hours, on the average :-(
Win10 64-bit; GTX 750Ti; driver 365.19

Profile skgiven
Volunteer moderator
Project tester
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,835,617,224
RAC: 310,037
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 43839 - Posted: 27 Jun 2016 | 8:04:46 UTC - in response to Message 43824.

Jim, thanks for this heads up. Switching to the 359.06 drivers has allowed one of my machines to make the 24 hour deadline for 7 of the last 11 WUs (sometimes by as little as 2.5 minutes). Previously it was missing all the 50% bonus times after "upgrading" to 362 drivers. I'm currently not having warm fuzzy feelings about nvidia.

You are welcome, and it surprised me a bit too. It was all due to someone on another forum though who noticed it first. I don't recall a comparable case for CUDA, though it used to happen all the time for the AMD OpenCl drivers.

The 6.12 app upgrade from 6.05 was similar, 5 1/2 years ago, and I recall successive driver 'updates' also slowed my GT240's down bit by bit.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Jim1348
Send message
Joined: 28 Jul 12
Posts: 462
Credit: 1,130,762,168
RAC: 13,533
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwat
Message 43841 - Posted: 27 Jun 2016 | 8:26:05 UTC - in response to Message 43839.

The 6.12 app upgrade from 6.05 was similar, 5 1/2 years ago, and I recall successive driver 'updates' also slowed my GT240's down bit by bit.

I have lost track of the world before Fermis.

By the way, the newer drivers may be better at OpenCl, at least on Folding. It is hard to get a direct comparison on my two GTX 980s, since I don't keep records and have to depend on the two cards getting the same type of work unit at the same time, which doesn't happen all that often. And if it is true on Folding, it may not be true on other OpenCl projects such as POEM; there is not as much consistency between projects as there is on CUDA that I have found. I think everyone would have to do their own tests to be sure.

Post to thread

Message boards : Graphics cards (GPUs) : CUDA 7.5 drivers are faster than CUDA 8 on GTX 970