Profiler counter "Dram reads/writes"

Accelerated Computing CUDA CUDA Programming and Performance
1

Hey there,

My computer has got both a GTX580 and a Tesla C2050. When i profile kernels on Compute Visual Profiler, i miss the “Dram reads” and “Dram writes” counter with the GTX580 whereas i got them with the Tesla C2050.

I’m using Windows 7 x64 with 266.58 whql driver.

It is normal not to get the acces to this counter on Geforce products ?

Thanks in advance.

2

I have the exact same problem. I have a 580 on 266.58, running windows 7 64-bit also. Any answer would be greatly appreciated.

3

Same here but a 570.

4

same here, with a 580, i cannot get memory throughput

using 260.19.26 driver for linux

and cuda 3.2 toolkit

edit: also, im having trouble getting the memory throutput for a 9800gtx+ because the profiler complains that needs the profile counters gld 32/63/128 which are for 1.2 and higher CC.

5

The “dram reads” and “dram writes” counters are not enabled in Visual Profiler for GTX 580 (and also some other 5xx series GeForce and Quadro cards). This issue is present in both CUDA 3.2 and CUDA 4.0 RC versions of the toolkit.

6

Thanks =)

Where did you get the info ? I have not seen anything on official docs, changelogs, forum or anything else…

7

thanks too.

i didnt know that.
now im in a little trouble, i need to get some performance values (flops and memory bandwidth) of a research kernel i have.
i have a 9800GTX and a GTX580 i have to get performance values for each one.

what ways do i have for measuring them??
thanks in advance

Cristobal