Wrong Visual Profiler Output for GTX285 Visual Profiler Output not coherent with Compute Capabilty

Accelerated Computing CUDA CUDA Programming and Performance
1

I have an ASUS ENGTX285 board and since I’m optimizing a code, running the openclprof (Visual Profiler for OpenCL) and then the Analyze Occupancy function, I got the following output:

Kernel details : Grid size: 360 x 1, Block size: 3 x 36 x 1

Register Ratio		= 0.75  ( 12288 / 16384 ) [14 registers per thread] 

Shared Memory Ratio	= 0.1875 ( 3072 / 16384 ) [468 bytes per Block] 

Active Blocks per SM	= 6 : 8

Active threads per SM	= 648 : 768

Occupancy		= 1  ( 24 / 24 )

Achieved occupancy 	= 1  (on 30 SMs)

Occupancy limiting factor = None

But, since the GTX285 compute capability is 1.3, shouldn’t the Visual Profiler output have these lines instead of the corresponding ones above?

Active threads per SM	= x : 1024

Occupancy		= 1  ( 32 / 32 )

I’m I wrong or the Visual Profiler is not correctly detecting the GTX285 card?

Or otherwise, is my code for some reason actually not using all the “graphic card capabilities”?

2

Problem solved!

I was running the Visual Profiler through a “ssh -X” connection, my computer have a graphics card with compute capability 1.1 and for some reason the Visual Profiler although run on the right machine with the GTX285 board, was calculating the occupancy using the compute capability information about my workstation.

3

Problem solved!

I was running the Visual Profiler through a “ssh -X” connection, my computer have a graphics card with compute capability 1.1 and for some reason the Visual Profiler although run on the right machine with the GTX285 board, was calculating the occupancy using the compute capability information about my workstation.