Archives Discussions

[Raistmer]

Does memory shared between 2 chips? Why second device could be slower?

AFAIK HD5970 has 2 separate GPU chips on card. Do these chips share common GPU memory pool?

I recived such report from one of testers, it runs app on 2 HD5970 GPUs:

"
The progress on the 4 tasks is thus much slower for two tasks (device 1 and 3)
At 46 minutes on card 1, gpu 0 is 9% done, and gpu 1 is 2,5% done.
At 41 minutes on card 2, gpu 0 is 8% done, and gpu 1 is 2,1% done.
"

What another possible reason for such behavior besides shared GPU memory?

[empty_knapsack]

Very good idea to check that 2nd GPU core in fact computes correct results. I've already post here topic about problems with 5970 -- under Windows and Catalyst 10.4 & 10.5 2nd core of 5970 just producing pure garbage also working slower than 1st core (and 1st core at least producing correct results).

I have no idea how it's possible to miss such bug, so I have feelings that 10.4 & 10.5 wasn't tested at all with 5970 by ATI. There no such problems with 10.2 & 10.3

[Raistmer]

Thanks for info!

ATI's people, any comments on topic, please?

[omkaranathan]

Multi-gpu issues are being looked-upon by developers. You can expect changes with one of the upcoming releases.

[dominik_g]

I don't have any problems with the second GPU on my HD 5970. I'm running Linux (Ubuntu 9.10) with Catalyst 10.3 and SDK 2.1

[Raistmer]

And I can confirm, that secondary chips on both HD5970 cards installed into that host produce incorrect results.
Also, primary chips on both cards produce correct results.
That is, problem definitely exists.
Hope it will be solved soon...

[Raistmer]

For now, what way of detection multi-GPU cards can be used to allow app to be executed only on first chip of each card?
(in CLInfo-based detection they looks like identical 4 different GPUs. How one can tell that there is 2 cards with 2 GPUs each? )

[nou]

from OpenCl info alone? i do not think you can. with aditional API from ATI you can. but why? it is bug. you should just provide for selection of GPU for user.