diff --git a/Makefile.cpu b/Makefile.cpu
index 0a7867f73..a64497ad7 100644
--- a/Makefile.cpu
+++ b/Makefile.cpu
@@ -22,12 +22,25 @@
 # This makefile will be extended/completed as we go.
 
 CC = g++
+OBJDIR = .obj.cpu
 
+# comment following and uncomment the next one to enable MKL library
+USE_ACML=1
+#USE_MKL=1
+
+ifdef USE_MKL
 MKL_PATH = /opt/intel/composer_xe_2013_sp1.0.080
-MKL_INCLUDE = $(MKL_PATH)/mkl/include
-MKL_LIB = -L$(MKL_PATH)/compiler/lib/intel64 -L$(MKL_PATH)/mkl/lib/intel64 -L$(MKL_PATH)/compiler/lib/mic -L$(MKL_PATH)/mkl/lib/mic -lmkl_intel_lp64 -lmkl_intel_thread -lmkl_core -lm -liomp5 -lpthread -lm
+MATHLIB_INCLUDE = $(MKL_PATH)/mkl/include
+MATHLIB_LIB = -L$(MKL_PATH)/compiler/lib/intel64 -L$(MKL_PATH)/mkl/lib/intel64 -L$(MKL_PATH)/compiler/lib/mic -L$(MKL_PATH)/mkl/lib/mic -lmkl_intel_lp64 -lmkl_intel_thread -lmkl_core -lm -liomp5 -lpthread
+MATHLIB_DEFINE = -DUSE_MKL
+else
+ACML_PATH = /usr/local/acml5.3.0/gfortran64
+MATHLIB_INCLUDE = $(ACML_PATH)/include
+MATHLIB_LIB = -L$(ACML_PATH)/lib -lacml -lm
+MATHLIB_DEFINE = -DUSE_ACML
+endif
 
-INCFLAGS = -I Common/Include -I Math/Math -I MachineLearning/cn -I $(MKL_INCLUDE)  
+INCFLAGS = -I Common/Include -I Math/Math -I MachineLearning/cn -I $(MATHLIB_INCLUDE)  
 
 COMMON_SRC = Common/fileutil.cpp Common/DataWriter.cpp Common/ConfigFile.cpp Common/DataReader.cpp \
              Common/Eval.cpp Common/File.cpp Common/BestGpu.cpp
@@ -54,13 +67,14 @@ CORE_SRC = $(CN_SRC) $(MATH_SRC) $(COMMON_SRC)
 
 SRC =  $(READER_SRC) $(CORE_SRC)
 
-CFLAGS = -std=c++0x -std=c++11 -DCPUONLY -D_POSIX_SOURCE -D_XOPEN_SOURCE=600 -D__USE_XOPEN2K -DUSE_MKL -fopenmp -fpermissive -fPIC
+CFLAGS = -std=c++0x -std=c++11 -DCPUONLY -D_POSIX_SOURCE -D_XOPEN_SOURCE=600 -D__USE_XOPEN2K $(MATHLIB_DEFINE) -fopenmp -fpermissive -fPIC
 
 DEBUG = -g
 
 all: cn.exe UCIFastReader SequenceReader LUSequenceReader
-	mkdir -p bin
-	mv cn.exe *.so bin/
+	mkdir -p bin.cpu
+	mv cn.exe *.so bin.cpu/
+	ln -sf bin.cpu bin
 
 UCIFastReader: ${UCIFASTREADER_SRC:.cpp=.o} ${CORE_SRC:.cpp=.o}
 	$(CC) $(DEBUG) -fPIC -shared -o $(addsuffix .so, $@) $^  
@@ -78,7 +92,7 @@ LUSequenceReader: ${LUSEQUENCEREADER_SRC:.cpp=.o} ${CORE_SRC:.cpp=.o}
 #	$(CC) -o $(addsuffix .so, $@) $^ -fPIC -shared 
 
 cn.exe: ${CORE_SRC:.cpp=.o}
-	$(CC) $(DEBUG) -o $@ $^ $(MKL_LIB) -fopenmp -ldl
+	$(CC) $(DEBUG) -o $@ $^ $(MATHLIB_LIB) -fopenmp -ldl
 
 -include ${SRC:.cpp=.d}
 
diff --git a/Makefile.gpu b/Makefile.gpu
index 911a60a6e..8eb72b014 100644
--- a/Makefile.gpu
+++ b/Makefile.gpu
@@ -24,15 +24,27 @@
 CC = g++
 NVCC = nvcc
 
+# comment following and uncomment the next one to enable MKL library
+USE_ACML=1
+#USE_MKL=1
+
+ifdef USE_MKL
+MKL_PATH = /opt/intel/composer_xe_2013_sp1.0.080
+MATHLIB_INCLUDE = $(MKL_PATH)/include
+MATHLIB_LIB = -L$(MKL_PATH)/compiler/lib/intel64 -L$(MKL_PATH)/mkl/lib/intel64 -L$(MKL_PATH)/compiler/lib/mic -L$(MKL_PATH)/mkl/lib/mic -lmkl_intel_lp64 -lmkl_intel_thread -lmkl_core -lm -liomp5 -lpthread
+MATHLIB_DEFINE = -DUSE_MKL
+else
+ACML_PATH = /usr/local/acml5.3.0/gfortran64
+MATHLIB_INCLUDE = $(ACML_PATH)/include
+MATHLIB_LIB = -L$(ACML_PATH)/lib -lacml -lm
+MATHLIB_DEFINE = -DUSE_ACML
+endif
+
 CUDA_PATH = /usr/local/cuda-6.5
 CUDA_INCLUDE = $(CUDA_PATH)/include
 CUDA_LIB = -L$(CUDA_PATH)/lib64 -lcublas -lcudart -lcuda -lcurand -lcusparse -lnvidia-ml
 
-MKL_PATH = /opt/intel/composer_xe_2013_sp1.0.080
-MKL_INCLUDE = $(MKL_PATH)/include
-MKL_LIB = -L$(MKL_PATH)/compiler/lib/intel64 -L$(MKL_PATH)/mkl/lib/intel64 -L$(MKL_PATH)/compiler/lib/mic -L$(MKL_PATH)/mkl/lib/mic -lmkl_intel_lp64 -lmkl_intel_thread -lmkl_core -lm -liomp5 -lpthread -lm
-
-INCFLAGS = -I Common/Include -I Math/Math -I MachineLearning/cn -I $(CUDA_INCLUDE) -I $(MKL_INCLUDE)
+INCFLAGS = -I Common/Include -I Math/Math -I MachineLearning/cn -I $(CUDA_INCLUDE) -I $(MATHLIB_INCLUDE)
 
 COMMON_SRC = Common/fileutil.cpp Common/DataWriter.cpp Common/ConfigFile.cpp Common/DataReader.cpp \
              Common/Eval.cpp Common/File.cpp Common/BestGpu.cpp
@@ -58,7 +70,7 @@ READER_SRC = $(UCIFASTREADER_SRC) $(LUSEQUENCEREADER_SRC) $(HTKMLFREADER_SRC) $(
 
 CORE_SRC = $(CN_SRC) $(MATH_SRC) $(COMMON_SRC)
 
-CFLAGS = -std=c++0x -std=c++11 -D_POSIX_SOURCE -D_XOPEN_SOURCE=600 -D__USE_XOPEN2K -DUSE_MKL -fopenmp -fpermissive 
+CFLAGS = -std=c++0x -std=c++11 -D_POSIX_SOURCE -D_XOPEN_SOURCE=600 -D__USE_XOPEN2K $(MATHLIB_DEFINE) -fopenmp -fpermissive 
 
 NVCCFLAGS = -std=c++11 -D_POSIX_SOURCE -D_XOPEN_SOURCE=600 -D__USE_XOPEN2K -arch=compute_20
 
@@ -67,8 +79,9 @@ DEBUG = -g
 GPU_DEBUG = -G
 
 all: cn.exe UCIFastReader SequenceReader LUSequenceReader
-	mkdir -p bin
-	mv cn.exe *.so bin/
+	mkdir -p bin.gpu
+	mv cn.exe *.so bin.gpu/
+	ln -sf bin.gpu bin
 
 tmp = ${CORE_SRC:.cpp=.o}
 
@@ -88,7 +101,7 @@ LUSequenceReader: ${LUSEQUENCEREADER_SRC:.cpp=.o} ${tmp:.cu=.o}
 #	$(CC) -o $(addsuffix .so, $@) $^ -fPIC -shared 
 
 cn.exe: ${tmp:.cu=.o}
-	$(CC) $(DEBUG) -o $@ $^ $(CUDA_LIB) $(MKL_LIB) -fopenmp -ldl -fPIC
+	$(CC) $(DEBUG) -o $@ $^ $(CUDA_LIB) $(MATHLIB_LIB) -fopenmp -ldl -fPIC 
 
 -include ${SRC:.cpp=.d}
 
diff --git a/Math/Math/CPUMatrix.cpp b/Math/Math/CPUMatrix.cpp
index 9fb349fbf..4e72902ba 100644
--- a/Math/Math/CPUMatrix.cpp
+++ b/Math/Math/CPUMatrix.cpp
@@ -3561,16 +3561,24 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         const char flag = 'A';
         if (sizeof(ElemType) == sizeof(double))
         {
+#ifndef USE_MKL
+            dgesvd('A', 'A', (int)m, (int)n, reinterpret_cast <double*>(A.m_pArray), (int)lda, reinterpret_cast <double*>(SIGMA.m_pArray), reinterpret_cast <double*>(U.m_pArray), (int)ldu, reinterpret_cast <double*>(VT.m_pArray), (int)ldvt, &info);
+#else
             //missing arguments fixed! --author Wengong Jin 2014/12/15
             double *work = new double[lwork];
             dgesvd(&flag, &flag, (int *) &m, (int *) &n, reinterpret_cast <double*>(A.m_pArray), (int *) &lda, reinterpret_cast <double*>(SIGMA.m_pArray), reinterpret_cast <double*>(U.m_pArray), (int *) &ldu, reinterpret_cast <double*>(VT.m_pArray), (int *) &ldvt, work, &lwork, &info);
+#endif
         }
         else
         {
 #pragma warning (suppress: 4244)
+#ifndef USE_MKL
+           sgesvd('A', 'A', (int)m, (int)n, reinterpret_cast <float*>(A.m_pArray), (int)lda, reinterpret_cast <float*>(SIGMA.m_pArray), reinterpret_cast <float*>(U.m_pArray), (int)ldu, reinterpret_cast <float*>(VT.m_pArray), (int)ldvt, &info);
+#else
             //missing arguments fixed! --author Wengong Jin 2014/12/15
             float *work = new float[lwork];
             sgesvd(&flag, &flag, (int *) &m, (int *) &n, reinterpret_cast <float*>(A.m_pArray), (int *) &lda, reinterpret_cast <float*>(SIGMA.m_pArray), reinterpret_cast <float*>(U.m_pArray), (int *) &ldu, reinterpret_cast <float*>(VT.m_pArray), (int *) &ldvt, work, &lwork, &info);
+#endif
         }
     }