diff --git a/benchmarks/Benchmark_dwf.cc b/benchmarks/Benchmark_dwf.cc
index 6da01b0d..2f258a48 100644
--- a/benchmarks/Benchmark_dwf.cc
+++ b/benchmarks/Benchmark_dwf.cc
@@ -170,12 +170,13 @@ int main (int argc, char ** argv)
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
 #endif
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
+  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGpu       ) std::cout << GridLogMessage<< "* Using Gpu WilsonKernels" <<std::endl;
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
   std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
 
   DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
-  int ncall =500;
+  int ncall =1000;
   if (1) {
     FGrid->Barrier();
     Dw.ZeroCounters();
@@ -201,7 +202,7 @@ int main (int argc, char ** argv)
     std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NN<<std::endl;
     err = ref-result; 
     std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
-
+    //    exit(0);
     /*
     if(( norm2(err)>1.0e-4) ) { 
       std::cout << "RESULT\n " << result<<std::endl;
@@ -256,6 +257,7 @@ int main (int argc, char ** argv)
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
 #endif
     if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
+    if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGpu       ) std::cout << GridLogMessage<< "* Using Gpu WilsonKernels" <<std::endl;
     if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
     if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
     std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
@@ -324,12 +326,10 @@ int main (int argc, char ** argv)
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
 #endif
-      if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) 
-	std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
-      if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) 
-	std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
-      if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) 
-	std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
+      if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   )  std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
+      if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGpu       )  std::cout << GridLogMessage<< "* Using Gpu WilsonKernels" <<std::endl;
+      if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll)  std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
+      if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm )  std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
       std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
 
       LatticeFermion sr_eo(sFGrid);
@@ -475,6 +475,7 @@ int main (int argc, char ** argv)
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
 #endif
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
+  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGpu       ) std::cout << GridLogMessage<< "* Using Gpu WilsonKernels" <<std::endl;
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
   std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
diff --git a/benchmarks/Benchmark_memory_bandwidth.cc b/benchmarks/Benchmark_memory_bandwidth.cc
index bc8ff76d..9b921a77 100644
--- a/benchmarks/Benchmark_memory_bandwidth.cc
+++ b/benchmarks/Benchmark_memory_bandwidth.cc
@@ -30,7 +30,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 using namespace std;
 using namespace Grid;
- ;
 
 int main (int argc, char ** argv)
 {
@@ -55,8 +54,8 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<"\t\t seconds"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
-  uint64_t lmax=96;
-#define NLOOP (10*lmax*lmax*lmax*lmax/vol)
+  uint64_t lmax=48;
+#define NLOOP (100*lmax*lmax*lmax*lmax/vol)
   for(int lat=8;lat<=lmax;lat+=8){
 
       Coordinate latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
@@ -73,14 +72,10 @@ int main (int argc, char ** argv)
       double a=2.0;
 
 
+      axpy(z,a,x,y);
       double start=usecond();
       for(int i=0;i<Nloop;i++){
-	auto x_v = x.View();
-	auto y_v = y.View();
-	auto z_v = z.View();
 	axpy(z,a,x,y);
-        x_v[0]=z_v[0]; // serial loop dependence to prevent optimise
-        y_v[4]=z_v[4];
       }
       double stop=usecond();
       double time = (stop-start)/Nloop*1000;
@@ -112,14 +107,10 @@ int main (int argc, char ** argv)
 
       uint64_t Nloop=NLOOP;
 
+      z=a*x-y;
       double start=usecond();
       for(int i=0;i<Nloop;i++){
-	auto x_v = x.View();
-	auto y_v = y.View();
-	auto z_v = z.View();
 	z=a*x-y;
-        x_v[0]=z_v[0]; // force serial dependency to prevent optimise away
-        y_v[4]=z_v[4];
       }
       double stop=usecond();
       double time = (stop-start)/Nloop*1000;
@@ -151,12 +142,10 @@ int main (int argc, char ** argv)
       LatticeVec y(&Grid);// random(pRNG,y);
       RealD a=2.0;
 
+      z=a*x;
       double start=usecond();
       for(int i=0;i<Nloop;i++){
-	auto x_v = x.View();
-	auto z_v = z.View();
 	z=a*x;
-        x_v[0]=z_v[0]*2.0;
       }
       double stop=usecond();
       double time = (stop-start)/Nloop*1000;
diff --git a/benchmarks/Benchmark_su3.cc b/benchmarks/Benchmark_su3.cc
index 61165888..f9f64b04 100644
--- a/benchmarks/Benchmark_su3.cc
+++ b/benchmarks/Benchmark_su3.cc
@@ -1,4 +1,4 @@
-    /*************************************************************************************
+/*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
 
@@ -31,16 +31,17 @@ Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
 using namespace std;
 using namespace Grid;
 
+
 int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
-#define LMAX (32)
-#define LMIN (4)
-#define LADD (4)
+#define LMAX (48)
+#define LMIN (8)
+#define LADD (8)
 
-  int64_t Nwarm=5;
-  int64_t Nloop=25;
+  int64_t Nwarm=50;
+  int64_t Nloop=500;
 
   Coordinate simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
   Coordinate mpi_layout  = GridDefaultMpi();
@@ -141,6 +142,7 @@ int main (int argc, char ** argv)
       double start=usecond();
       for(int64_t i=0;i<Nloop;i++){
 	mult(z,x,y);
+	//	mac(z,x,y);
       }
       double stop=usecond();
       double time = (stop-start)/Nloop*1000.0;
@@ -179,7 +181,7 @@ int main (int argc, char ** argv)
       double stop=usecond();
       double time = (stop-start)/Nloop*1000.0;
       
-      double bytes=3*vol*Nc*Nc*sizeof(Complex);
+      double bytes=4*vol*Nc*Nc*sizeof(Complex);
       double flops=Nc*Nc*(8+8+8)*vol;
       std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<"   \t\t"<<bytes/time<<"\t\t" << flops/time<<std::endl;