This is for testing the new application on older cuda.

We will keep only cuda4.0 and cuda4.2 applications. The second one is required only if you have a kepler card and it needs newest drivers. So, most people will be actually using cuda4.0 which is being tested now in beta.

If you have a cuda4.2 driver, the server should give you that. If you have a older driver but more recent than 270.41, then you should receive cuda4.0.

Another change is that as most of valuable work in the long queue, we will be testing the new application on the acemd2 queue as soon as the beta ends. Only later extend it to acemdlong.

Is it a problem to have only cuda4.2 for Linux?

gdf

gdf

Is it a problem to have only cuda4.2 for Linux?

No problem for us.

Hi, The latest drivers NVIDIA for Linux (Ubuntu) already support CUDA 4.2. Greetings.

GTX560Ti CC2.1 XP 64bit driver 301.24, core 6.10.58: 3 tasks errored out at beginning
GT555M W7 64bit driver 301.42, core 7.0.25: 1 task errored out after 60s
Both machines changed profile (no beta more).

This windows beta is giving major screen lag in XP 64. Driver 301.42, 7.0.25, GTX580. Have suspended while I still can. One completed OK & one failed after 20secs.
Update: Tried Boinc 7.0.28 but still the same lag, making computer unusable. Tasks finished OK though.

This is for testing the new application on older cuda.

We will keep only cuda4.0 and cuda4.2 applications. The second one is required only if you have a kepler card and it needs newest drivers. So, most people will be actually using cuda4.0 which is being tested now in beta.

If you have a cuda4.2 driver, the server should give you that. If you have a older driver but more recent than 270.41, then you should receive cuda4.0.

Another change is that as most of valuable work in the long queue, we will be testing the new application on the acemd2 queue as soon as the beta ends. Only later extend it to acemdlong.

Is it a problem to have only cuda4.2 for Linux?

gdf

gdf

The priority needs to be set to below normal, because these units make the computer slow and sluggish.

The cuda 4.0 units are about 50+ seconds slower than the cuba 4.2 units on my computers, so why bother with cuba 4.0 all together, and have everyone update their drivers?

On my 680 I'm receiving both CUDA 4 and 4.2. Mainly 4.2, and as expected the 4.0 error out at 0.00

Driver 302.59

Assuming the cuda40 and 42 beta work units are the same size the cuda42 seems to be the fastest on cpu and run-time on my GTX570. Link to one host here

I would suggest cuda42 be the default for Fermi and Kepler based cards (and appropiate driver version), otherwise cuda40 as the fallback.

Also machine response is very bad when running these, presumable because of their priority as other have mentioned.

---

BOINC blog

A fix coming up tomorrow. Now you should be getting always the same application.
Either 4.2 or 4.0 depending on the driver. Not a mixture of the two...

gdf

The cuda 4.0 units are about 50+ seconds slower than the cuba 4.2 units on my computers, so why bother with cuba 4.0 all together, and have everyone update their drivers?

The project is trying to be nice to crunchers. For some updating the driver is a nightmare or not [easily] possible due to various reasons.

MrS

---

Scanning for our furry friends since Jan 2002

From discussions with the Seti optimisers (well one of them anyway), he said that the older GTX2xx cards are worse off running cuda42 apps. It would seem cuda42 is particularly optimised for Fermi and Kepler chips. The old cards can still run them, they are slower than a cuda32 app on the same card. I don't think they ever tried a cuda40 app, but they did try cuda41 which also was worse than cuda32 on older cards.

You might need to consider a cuda32 app for the older cards, even if the user has up to date drivers. That is a cuda32 instead of a cuda40 app. You could use the compute capability to determine if they should get cuda32 app rather than driver version. Something like (if compute capability = 1.3 then use cuda32)

---

BOINC blog

"nenym" wrote:

GTX560Ti CC2.1 XP 64bit driver 301.24, core 6.10.58: 3 tasks errored out at beginning

The same one beta CUDA 42 task with 301.42 driver.

The new app does not compile on anything older than cuda4.0.

gdf

From discussions with the Seti optimisers (well one of them anyway), he said that the older GTX2xx cards are worse off running cuda42 apps. It would seem cuda42 is particularly optimised for Fermi and Kepler chips. The old cards can still run them, they are slower than a cuda32 app on the same card. I don't think they ever tried a cuda40 app, but they did try cuda41 which also was worse than cuda32 on older cards.

You might need to consider a cuda32 app for the older cards, even if the user has up to date drivers. That is a cuda32 instead of a cuda40 app. You could use the compute capability to determine if they should get cuda32 app rather than driver version. Something like (if compute capability = 1.3 then use cuda32)

The new app does not compile on anything older than cuda4.0.

gdf

From discussions with the Seti optimisers (well one of them anyway), he said that the older GTX2xx cards are worse off running cuda42 apps. It would seem cuda42 is particularly optimised for Fermi and Kepler chips. The old cards can still run them, they are slower than a cuda32 app on the same card. I don't think they ever tried a cuda40 app, but they did try cuda41 which also was worse than cuda32 on older cards.

You might need to consider a cuda32 app for the older cards, even if the user has up to date drivers. That is a cuda32 instead of a cuda40 app. You could use the compute capability to determine if they should get cuda32 app rather than driver version. Something like (if compute capability = 1.3 then use cuda32)

If cuda40 is the minimum version you can do, can you supply it based on the compute capability of the card?

---

BOINC blog

New cuda4.0 app is out now for windows only.

Let us know.

gdf

The first task (cuda40) took 2 minutes & all was fine. The other tasks (cuda42) are causing the screen to become unresponsive. Tasks finishing without error. XP64,GTX580, 7.0.28, 301.42.
Getting more cuda40 now. Guess these are all configured differently? This one took 16m30s

Welcome to the 2D clock bug

GPU is at 50MHz

Yeah. Just rebooted & back to normal.

Is yours a factory overclocked card?

With these newer drivers, there are quite a few people on NVIDIA's forums with factory OC, which are getting these types of issues.

Will have to see if this keeps popping up I guess with everyone else.

Cheers

I flashed the bios myself a year ago & never had any issues with it. I think the driver just crashed rather than down clocked. That slow cuda40 task was straight after a cuda42 one which was causing lag.