I've upgraded my old 8600GT to 9600GT (PCIe x16 slot in the motherboard) and moved my old 8600GT to eht 2nd PCIe slot (x16 physical slot, x4 electrical wiring).

All seems to work fine except that BOINC 6.4.5 (running on Win XP SP3 box) reports that is has found 8600GT (2) instead of 9600GT (1) and 8600GT (1).

Is this a cosmetic bug or a detection issue? And does it have an impact on work unit results?

Thanks for your input

try http://setiathome.berkeley.edu/forum_thread.php?id=50997&nowrap=true#846235

Atm. i would say it isnt a good idea to run cuda with 2 cards.
They have much trouble to get them running on one card so i think they will fix that first and then the 'two cards' - problem ...

---

So apparently it is detected correctly but the aggregated summary isn't correct.

From one of the uploaded work units:
<core_client_version>6.4.5</core_client_version>
<![CDATA[
<stderr_txt>
cudaAcc_initializeDevice: Found 2 CUDA device(s):
Device 1 : GeForce 9600 GT
Device 2 : GeForce 8600 GT
cudaAcc_initializeDevice is determiming what CUDA device to use...
user specified SETI to use CUDA device 1: GeForce 9600 GT
SETI@home using CUDA accelerated device GeForce 9600 GT
setiathome_enhanced 6.02 Visual Studio/Microsoft C++
libboinc: 6.3.22

And for another one:
<core_client_version>6.4.5</core_client_version>
<![CDATA[
<stderr_txt>
cudaAcc_initializeDevice: Found 2 CUDA device(s):
Device 1 : GeForce 9600 GT
Device 2 : GeForce 8600 GT
cudaAcc_initializeDevice is determiming what CUDA device to use...
user specified SETI to use CUDA device 2: GeForce 8600 GT
SETI@home using CUDA accelerated device GeForce 8600 GT
setiathome_enhanced 6.02 Visual Studio/Microsoft C++
libboinc: 6.3.22

But the computer summary and the GUI of the client show:
[CUDA|GeForce 8600 GT|2|511MB]

So I suspect this is purely cosmetic and will get fixed in some futher release

---

So apparently it is detected correctly but the aggregated summary isn't correct.

This has an annoying side effect, too. If your slowest card cannont run CUDA apps (for example because it has not enough RAM on board), neither the other one will run them.

This prevented me from running CUDA when I first tried it: my on-board 8300 was reported as having some 250 MB of RAM, which is less than the minimum required (this should be changed, BTW: if 256 megs are required, this should mean "around 256", not "exactly 256"). Since both my card where reported to S&H as 8300s, I could not get CUDA apps.

When I assigned 512 megs to the onboard card I was able to run CUDA on both cards.

Regards,

Roberto

---

BTW: if 256 megs are required, this should mean "around 256", not "exactly 256").

It's 254MB, as most cards detected with 256MB report they have 255MB of memory.
If it needs to be changed any further, then where do the developers stop? 250MB and getting complaints that 248MB isn't detected?

---

If it needs to be changed any further, then where do the developers stop? 250MB and getting complaints that 248MB isn't detected?

I see your point. I suppose no 256-megs card is actually reported far from 250MB, so you might lower the limit to 240 megs and be sure that almost all boards fall within it.

It's of course up to you pepole to decide whether it is better to keep that limit as close as possible to 256 MB (for example because the CUDA apps really needs that amount of RAM :) ), or to lower it a bit to include as many boards as possible and thus rise the total computing power available to S&H.

Regards,

Roberto

---

It's of course up to you people to decide whether it is better to keep that limit as close as possible to 256 MB (for example because the CUDA apps really needs that amount of RAM :) ), or to lower it a bit to include as many boards as possible and thus rise the total computing power available to S&H.

The CUDA application uses a minimum of 200MB of the GPUs RAM at all times, per task. It can spike upwards when in use.

The change from 256MB to 254MB was only made to make sure those cards whose memory is 256MB and were detected as 255MB were included.

I'm sure the developers want to fix the abundance of errors thrown out by their application first before considering to lower the minimal RAM needed on a card.

---