require sm_2x and above

diff --git a/src/Makefile b/src/Makefile
index bbbfd27..60aa139 100644 (file)
--- a/src/Makefile
+++ b/src/Makefile
@@ -31,7 +31,7 @@ TEST_BINS := ${TEST_OBJS:.o=.testbin}
 
 # define third-party library paths
 CUDA_DIR := /usr/local/cuda
-CUDA_ARCH := -arch=sm_20
+CUDA_ARCH := -arch=sm_30
 MKL_DIR := /opt/intel/mkl
 
 CUDA_INCLUDE_DIR := $(CUDA_DIR)/include

diff --git a/src/caffe/common.hpp b/src/caffe/common.hpp
index c28ad57..8eb7987 100644 (file)
--- a/src/caffe/common.hpp
+++ b/src/caffe/common.hpp
@@ -49,8 +49,8 @@ namespace caffe {
 using boost::shared_ptr;
 
 
-// For backward compatibility we will just use 512 threads per block
-const int CAFFE_CUDA_NUM_THREADS = 512;
+// We will use 1024 threads per block, which requires cuda sm_2x or above.
+const int CAFFE_CUDA_NUM_THREADS = 1024;
 
 
 inline int CAFFE_GET_BLOCKS(const int N) {