Do different types of GPUGrid tasks use different amounts of GPU RAM, especially at startup?

I have two 512MB 9800GT-class cards, both testing the new v6.10.46 BOINC client. Both are running ACEMD v6.03 tasks.

One is running p47-IBUCH_0510_pYEEI_long_100421-0-12-RND3168_0 without problems.

The other tried to start f237-TONI_CAPBIND99SB-51-100-RND2375_1, but failed with

22-Apr-2010 11:02:12 [GPUGRID] [coproc_debug] CUDA device 0 lacks RAM for f237-TONI_CAPBIND99SB-51-100-RND2375_1

That seems to be specific to v6.10.46: I downgraded to v6.10.45, and the task is now running normally.

The <app_version> for v6.03 says

<gpu_ram>249561088.000000</gpu_ram>

- 512 MB should be enough, surely?

I think I can duplicate the problem on Vista64, but not on Windows XP. I do not have that debug message about memory, probably because I do not have debug turned on.

Here is what I found on Vista-64. I first enabled gpugrid for two tasks and got only pair of IBUCH for 6.10.43. This seemed to work and I then stopped boinc and installed 6.10.46. I have a gtx9800 (512mb) and a gts250 (1024mb)

The two tasks started up ok.

2010-04-22 10:33:11 AM NVIDIA GPU 0: GeForce 9800 GTX/9800 GTX+ (driver version 19745, CUDA version 3000, compute capability 1.1, 496MB, 442 GFLOPS peak)
2010-04-22 10:33:11 AM NVIDIA GPU 1: GeForce GTS 250 (driver version 19745, CUDA version 3000, compute capability 1.1, 1006MB, 470 GFLOPS peak)
2010-04-22 10:33:12 AM GPUGRID Restarting task p34-IBUCH_51_pYEEI_100301-26-40-RND8331_1 using acemd2 version 603
2010-04-22 10:33:12 AM GPUGRID Restarting task p32-IBUCH_0511_pYEEI_long_100421-1-12-RND6657_1 using acemd2 version 603

<cc_config>
<options>
<use_all_gpus>1</use_all_gpus>
</options>
</cc_config>

but within about 30 seconds the one running on the GTX9800 stopped and started waiting to run. Note that I have enabled the use_all_gpus as shown above. I failed to copy the "before" but it might still be on the disk. During the 30 seconds or so it was running, it was going so slow that I knew somthing was wrong. While I was watching the second tick off I saw the message about waiting show up and it stopped crunching.




===then I tried same thing on windows xp that has two identical gtx9800+


Second, Windows XP (32 bit) 6.10.43 ran ok as follows:

I run the following on pair of 9800gtx+

jys-xp-2duo

5232 GPUGRID 4/22/2010 9:49:33 AM Starting p46-IBUCH_101c_pYEEI_100304-26-80-RND3796_0
5233 GPUGRID 4/22/2010 9:49:34 AM Starting task p46-IBUCH_101c_pYEEI_100304-26-80-RND3796_0 using acemd2 version 603
5234 GPUGRID 4/22/2010 9:49:34 AM Starting g125-TONI_CAPBIND99SB-58-100-RND5760_0
5235 GPUGRID 4/22/2010 9:49:35 AM Starting task g125-TONI_CAPBIND99SB-58-100-RND5760_0 using acemd2 version 603

I closed boinc and restarted and the tasks picked up ok. I then installed in 6.10.46 and it picked up ok. I then stopped it and restarted and it still ran ok.

1 4/22/2010 10:18:28 AM Starting BOINC client version 6.10.46 for windows_intelx86
jys-xp-2duo

15 4/22/2010 10:18:28 AM NVIDIA GPU 0: GeForce 9800 GTX/9800 GTX+ (driver version 19713, CUDA version 3000, compute capability 1.1, 512MB, 470 GFLOPS peak)
16 4/22/2010 10:18:28 AM NVIDIA GPU 1: GeForce 9800 GTX/9800 GTX+ (driver version 19713, CUDA version 3000, compute capability 1.1, 512MB, 470 GFLOPS peak)
46 GPUGRID 4/22/2010 10:18:29 AM Restarting task p46-IBUCH_101c_pYEEI_100304-26-80-RND3796_0 using acemd2 version 603
47 GPUGRID 4/22/2010 10:18:29 AM Restarting task g125-TONI_CAPBIND99SB-58-100-RND5760_0 using acemd2 version 603

[EDIT] Not working for windows xp either. Just takes a little longer, like 2 minutes and CAPBIND starts waiting. Then, it gets resumed and later it waits agin.

Mine is Windows XP, 32-bit, and it's continuing under BOINC v6.10.47:

22-Apr-2010 14:48:00 [SETI@home] Computation for task 28dc06aa.29060.3753.5.10.150_1 finished
22-Apr-2010 14:48:00 [GPUGRID] [coproc_debug] Assigning CUDA instance 0 to f237-TONI_CAPBIND99SB-51-100-RND2375_1
22-Apr-2010 14:48:00 [GPUGRID] [cpu_sched] Starting f237-TONI_CAPBIND99SB-51-100-RND2375_1(resume)
22-Apr-2010 14:48:00 [GPUGRID] Restarting task f237-TONI_CAPBIND99SB-51-100-RND2375_1 using acemd2 version 603
22-Apr-2010 14:48:03 [GPUGRID] [coproc_debug] CUDA device 0 lacks RAM for f237-TONI_CAPBIND99SB-51-100-RND2375_1
22-Apr-2010 14:48:03 [GPUGRID] [cpu_sched] Preempting f237-TONI_CAPBIND99SB-51-100-RND2375_1 (removed from memory)
22-Apr-2010 14:48:05 [GPUGRID] Starting p9-IBUCH_1010_PQpYIpYIPI_long_100421-1-4-RND6327_1
22-Apr-2010 14:48:05 [GPUGRID] [cpu_sched] Starting p9-IBUCH_1010_PQpYIpYIPI_long_100421-1-4-RND6327_1 (initial)
22-Apr-2010 14:48:05 [GPUGRID] Starting task p9-IBUCH_1010_PQpYIpYIPI_long_100421-1-4-RND6327_1 using acemd2 version 603

and away it ran.

Any ideas from Project Admin or developers on this?

Mine is Windows XP, 32-bit, and it's continuing under BOINC v6.10.47:

22-Apr-2010 14:48:00 [SETI@home] Computation for task 28dc06aa.29060.3753.5.10.150_1 finished
22-Apr-2010 14:48:00 [GPUGRID] [coproc_debug] Assigning CUDA instance 0 to f237-TONI_CAPBIND99SB-51-100-RND2375_1
22-Apr-2010 14:48:00 [GPUGRID] [cpu_sched] Starting f237-TONI_CAPBIND99SB-51-100-RND2375_1(resume)
22-Apr-2010 14:48:00 [GPUGRID] Restarting task f237-TONI_CAPBIND99SB-51-100-RND2375_1 using acemd2 version 603
22-Apr-2010 14:48:03 [GPUGRID] [coproc_debug] CUDA device 0 lacks RAM for f237-TONI_CAPBIND99SB-51-100-RND2375_1
22-Apr-2010 14:48:03 [GPUGRID] [cpu_sched] Preempting f237-TONI_CAPBIND99SB-51-100-RND2375_1 (removed from memory)
22-Apr-2010 14:48:05 [GPUGRID] Starting p9-IBUCH_1010_PQpYIpYIPI_long_100421-1-4-RND6327_1
22-Apr-2010 14:48:05 [GPUGRID] [cpu_sched] Starting p9-IBUCH_1010_PQpYIpYIPI_long_100421-1-4-RND6327_1 (initial)
22-Apr-2010 14:48:05 [GPUGRID] Starting task p9-IBUCH_1010_PQpYIpYIPI_long_100421-1-4-RND6327_1 using acemd2 version 603

and away it ran.

Any ideas from Project Admin or developers on this?

I just checked my Windows XP-32 bit and indeed, it has the same problem. Just takes longer to happen and it can resume all by itself and then after a few minutes starts waiting agin.

Setiathome is failing also: (vista 64 with gts250(1024) and 9800gtx(512)

the 512mb card allows seti to start, it then starts waiting, and another seti is then allowed to start which then waits. Exact same thing on windows xp with pair of 9800gtx+

collatz seems to run fine on both cards and on both xp-32 and vista64.