Does anyone have a method for determining the number of memory partitions on a GPU? A definitive list of possible values for cudaDeviceProp.name would probably be sufficient. I know there is a list in the programming guide but it doesn’t include the “Tesla T10 Processor” for example. I’m primarily interested in GT200 series desktop and mobile GPUs.
For those DDR3 era GPUs (so G80/G90 and GT200), I think you can device the memory bus pin count by 64 to get the number of banks.
I think so too, but how do I get the memory bus pin count?
Not available from the API, you have to hit the specifications for that. For the GT200 cards it was pretty simple IIRC: 8 banks for the GTX280/285 and Tesla T10 family, 7 banks for the GTX295/275/260. For the G80/G92 the Tesla C870 and relatives were 6 banks, the consumer G80/G92 parts were 4 banks.
Yes, that sounds about right. But how can my software identify what card its running on? This is what I need a definitive list of possible values for cudaDeviceProp.name for. Or is “Tesla T10 Processor” perhaps the only omission from the list in the programming guide?
Edit: Its probably enough for my application to identify the GTX 295/275/260 as 7 banks, the Quadro FX 4800 as 6 banks and assume everything else is 8 banks for now. I might need to do something different for G80/G92 and Fermi later.