How many "CUDA cores" does each GPU multiprocessor use?

Question

I know that devices prior to Fermi architecture had 8 SPs in one multiprocessor. Is the graph the same in Fermi architecture?

Answer 1

The number of Multiprocessors (MP) and the number of cores per MP can be found by running DeviceQuery.exe . It is located in the %NVSDKCOMPUTE_ROOT%/C/bin the SDK Computing SDK installation.

Look at the DeviceQuery code (found in %NVSDKCOMPUTE_ROOT%/C/src/DeviceQuery ), which shows that the number of cores is calculated by passing the xy CUDA Capability values ​​to the ConvertSMVer2Cores utility function.

From the ConvertSMVer2Cores code, you can see this relationship between features and the number of cores:

 Capability: Cores 10: 8 11: 8 12: 8 13: 8 20: 32 21: 48 

Answer 2

The answer depends on the Compute Capability property of the CUDA device. Figures:

• Computing ability <= 1.3 → 8 CUDA Cores / SM
• CC == 2.0 → 32 CUDA / SM cores
• CC == 2.1 → 48 CUDA core / SM

See Appendix G CUDA C Programming Guide .

Answer 3

Updating @AshwinNanjappa answer for CUDA 7.5:

 Compute Capability # Cores ----------------------- 1.x: 8 2.0: 32 2.1: 48 3.x: 192 5.x: 128 

Notes:

• CUDA 7.5 no longer supports a device with computing power of 1.x.
• In fact, these are not “cores” in the sense of processors. See this question here in the Stack Overflow section.
• In Maxwell GPU (5.x) the number of "cores" per multiprocessor is reduced.
• I got additional information from $CUDA_SAMPLES_DIR/common/inc/helper_cuda.h .