SM 3.2 specs


AFAIK the TK1 (GK20A arch.) has compute capability 3.2 which is said to be “roughly the same” as CC 3.5.

Anyone know where I might find the exact specs on this?

I know the SM differs in that the register file is smaller (half the size).

-Any other known differences?
-Has anyone tested atomic performance?

/deviceQuery Starting...

CUDA Device Query (Runtime API) version (CUDART static linking)

Detected 1 CUDA Capable device(s)

Device 0: "GK20A"
CUDA Driver Version / Runtime Version 6.0 / 6.0
CUDA Capability Major/Minor version number: 3.2
Total amount of global memory: 1746 MBytes (1831051264 bytes)
( 1) Multiprocessors, (192) CUDA Cores/MP: 192 CUDA Cores
GPU Clock rate: 960 MHz (0.96 GHz)
Memory Clock rate: 924 Mhz
Memory Bus Width: 64-bit
L2 Cache Size: 131072 bytes
Maximum Texture Dimension Size (x,y,z) 1D=(65536), 2D=(65536, 65536), 3D=(4096, 4096, 4096)
Maximum Layered 1D Texture Size, (num) layers 1D=(16384), 2048 layers
Maximum Layered 2D Texture Size, (num) layers 2D=(16384, 16384), 2048 layers
Total amount of constant memory: 65536 bytes
Total amount of shared memory per block: 49152 bytes
Total number of registers available per block: 32768
Warp size: 32
Maximum number of threads per multiprocessor: 2048
Maximum number of threads per block: 1024
Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
Max dimension size of a grid size (x,y,z): (2147483647, 65535, 65535)
Maximum memory pitch: 2147483647 bytes
Texture alignment: 512 bytes
Concurrent copy and kernel execution: Yes with 1 copy engine(s)
Run time limit on kernels: No
Integrated GPU sharing Host Memory: Yes
Support host page-locked memory mapping: No
Alignment requirement for Surfaces: Yes
Device has ECC support: Disabled
Device supports Unified Addressing (UVA): Yes
Device PCI Bus ID / PCI location ID: 0 / 0
Compute Mode:
< Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >

Some microbenchmarks are definitely needed.

The only sm_32 vs. sm_35 SMX test I’ve done appears to indicate shared memory bandwidth has been halved but that might be entirely due to launching different size grids for each arch.

The GK20A produces plenty of nvprof statistics. The ones that stood out in my quick test of a single block running on a GK20A vs. GK208 are:

• IPC: 1.33 vs. 2.0
• Shared load: 23 vs. 43 GB/sec
• Shared store 21 vs. 38 GB/sec

Maybe someone has a suite of microbenchmarks that they would like to run on the GK20A?