diff --git a/Grid/communicator/Communicator_mpi3.cc b/Grid/communicator/Communicator_mpi3.cc
index 6b6c9dec..7dc706df 100644
--- a/Grid/communicator/Communicator_mpi3.cc
+++ b/Grid/communicator/Communicator_mpi3.cc
@@ -408,8 +408,7 @@ double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequ
 {
   return 0.0; // Do nothing -- no preparation required
 }
-double CartesianCommunicator::StencilSendToRecvFromBegin(int list_idx,
-							 std::vector<CommsRequest_t> &list,
+double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
 							 int dest,int dox,
 							 void *recv,
@@ -470,6 +469,7 @@ void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsReque
   int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
   assert(ierr==0);
   list.resize(0);
+  this->StencilBarrier(); 
 }
 
 #else /* NOT     ... ACCELERATOR_AWARE_MPI */
@@ -481,10 +481,10 @@ void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsReque
    * PHASE 1: (prepare)
    * - post MPI receive buffers asynch
    * - post device - host send buffer transfer asynch
-   * - post device - device transfers
    * PHASE 2: (Begin)
    * - complete all copies
    * - post MPI send asynch
+   * - post device - device transfers
    * PHASE 3: (Complete)
    * - MPI_waitall
    * - host-device transfers
@@ -561,6 +561,8 @@ double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequ
   
   if (dox) {
     if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
+#undef DEVICE_TO_HOST_CONCURRENT // pipeline
+#ifdef DEVICE_TO_HOST_CONCURRENT
       tag= dir+_processor*32;
 
       host_xmit = this->HostBufferMalloc(xbytes);
@@ -577,11 +579,30 @@ double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequ
       srq.host_buf   = host_xmit;
       srq.device_buf = xmit;
       list.push_back(srq);
+#else
+      tag= dir+_processor*32;
 
-    } else {
-      void *shm = (void *) this->ShmBufferTranslate(dest,recv);
-      assert(shm!=NULL);
-      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
+      host_xmit = this->HostBufferMalloc(xbytes);
+      const int chunks=1;
+      for(int n=0;n<chunks;n++){
+	void * host_xmitc = (void *)( (uint64_t) host_xmit + n*xbytes/chunks);
+	void * xmitc      = (void *)( (uint64_t) xmit      + n*xbytes/chunks);
+	acceleratorCopyFromDeviceAsynch(xmitc, host_xmitc,xbytes/chunks); // Make this Asynch
+      }
+      acceleratorCopySynchronise(); // Complete all pending copy transfers
+      
+      ierr =MPI_Isend(host_xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+      assert(ierr==0);
+      off_node_bytes+=xbytes;
+
+      CommsRequest_t srq;
+      srq.PacketType = InterNodeXmit;
+      srq.bytes      = xbytes;
+      srq.req        = xrq;
+      srq.host_buf   = host_xmit;
+      srq.device_buf = xmit;
+      list.push_back(srq);
+#endif
     }
   }
 
@@ -623,17 +644,17 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
    * - complete all copies
    * - post MPI send asynch
    */
-  acceleratorCopySynchronise();
 
-  static int printed;
-  if(!printed && this->IsBoss() ) {
-    printf("dir %d doX %d doR %d Face size %ld %ld\n",dir,dox,dor,xbytes,rbytes);
-    printed=1;
-  }
+  //  static int printed;
+  //  if((printed<8) && this->IsBoss() ) {
+  //    printf("dir %d doX %d doR %d Face size %ld %ld\n",dir,dox,dor,xbytes,rbytes);
+  //    printed++;
+  //  }
   
   if (dox) {
 
     if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
+#ifdef DEVICE_TO_HOST_CONCURRENT
       tag= dir+_processor*32;
       // Find the send in the prepared list
       int list_idx=-1;
@@ -652,7 +673,12 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
       assert(ierr==0);
       list[list_idx].req        = xrq; // Update the MPI request in the list
       off_node_bytes+=xbytes;
-    } 
+#endif      
+    } else {
+      void *shm = (void *) this->ShmBufferTranslate(dest,recv);
+      assert(shm!=NULL);
+      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
+    }
   }
   return off_node_bytes;
 }
@@ -680,6 +706,7 @@ void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsReque
   acceleratorCopySynchronise(); // Complete all pending copy transfers
   list.resize(0);               // Delete the list
   this->HostBufferFreeAll();    // Clean up the buffer allocs
+  this->StencilBarrier(); 
 }
 #endif
 ////////////////////////////////////////////
diff --git a/Grid/communicator/SharedMemoryMPI.cc b/Grid/communicator/SharedMemoryMPI.cc
index ce11714f..dc22aee0 100644
--- a/Grid/communicator/SharedMemoryMPI.cc
+++ b/Grid/communicator/SharedMemoryMPI.cc
@@ -543,6 +543,9 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
 #ifndef ACCELERATOR_AWARE_MPI
   printf("Host buffer allocate for GPU non-aware MPI\n");
+#if 0
+  HostCommBuf= acceleratorAllocHost(bytes);
+#else 
   HostCommBuf= malloc(bytes); /// CHANGE THIS TO malloc_host
 #ifdef HAVE_NUMAIF_H
   #warning "Moving host buffers to specific NUMA domain"
@@ -569,6 +572,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   }
 #endif  
   acceleratorPin(HostCommBuf,bytes);
+#endif  
+
 #endif  
   ShmCommBuf = acceleratorAllocDevice(bytes);
   if (ShmCommBuf == (void *)NULL ) {
diff --git a/Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h b/Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h
index 14132cef..3d4e5cc5 100644
--- a/Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h
+++ b/Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h
@@ -332,7 +332,8 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st,
   
   //  std::cout << " WilsonFermion5D Communicate Begin " <<std::endl;
   std::vector<std::vector<CommsRequest_t> > requests;
-#ifndef GRID_ACCELERATED
+
+#if 1
   /////////////////////////////
   // Overlap with comms
   /////////////////////////////
@@ -352,7 +353,8 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st,
     Kernels::DhopKernel   (Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,1,0);
   }
   
-#ifdef GRID_ACCELERATED
+  //ifdef GRID_ACCELERATED
+#if 0
   /////////////////////////////
   // Overlap with comms -- on GPU the interior kernel call is nonblocking
   /////////////////////////////
diff --git a/Grid/stencil/Stencil.h b/Grid/stencil/Stencil.h
index 2de50e9c..1142891a 100644
--- a/Grid/stencil/Stencil.h
+++ b/Grid/stencil/Stencil.h
@@ -376,6 +376,7 @@ public:
 					  Packets[i].from_rank,Packets[i].do_recv,
 					  Packets[i].xbytes,Packets[i].rbytes,i);
     }
+    acceleratorCopySynchronise();
     for(int i=0;i<Packets.size();i++){
       _grid->StencilSendToRecvFromBegin(MpiReqs,
 					Packets[i].send_buf,
@@ -401,7 +402,6 @@ public:
     else DslashLogFull();
     //    acceleratorCopySynchronise();// is in the StencilSendToRecvFromComplete
     //    accelerator_barrier(); 
-    _grid->StencilBarrier(); 
     for(int i=0;i<Packets.size();i++){
       if ( Packets[i].do_recv )
 	FlightRecorder::recvLog(Packets[i].recv_buf,Packets[i].rbytes,Packets[i].from_rank);
diff --git a/Grid/threads/Accelerator.h b/Grid/threads/Accelerator.h
index 2862d087..c28ca201 100644
--- a/Grid/threads/Accelerator.h
+++ b/Grid/threads/Accelerator.h
@@ -209,6 +209,17 @@ void Lambda6Apply(uint64_t num1, uint64_t num2, uint64_t num3,
     }									\
   }
 
+inline void *acceleratorAllocHost(size_t bytes)
+{
+  void *ptr=NULL;
+  auto err = cudaMallocHost((void **)&ptr,bytes);
+  if( err != cudaSuccess ) {
+    ptr = (void *) NULL;
+    printf(" cudaMallocHost failed for %d %s \n",bytes,cudaGetErrorString(err));
+    assert(0);
+  }
+  return ptr;
+}
 inline void *acceleratorAllocShared(size_t bytes)
 {
   void *ptr=NULL;
@@ -230,8 +241,10 @@ inline void *acceleratorAllocDevice(size_t bytes)
   }
   return ptr;
 };
+
 inline void acceleratorFreeShared(void *ptr){ cudaFree(ptr);};
 inline void acceleratorFreeDevice(void *ptr){ cudaFree(ptr);};
+inline void acceleratorFreeHost(void *ptr){ cudaFree(ptr);};
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { cudaMemcpy(to,from,bytes, cudaMemcpyHostToDevice);}
 inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);}
 inline void acceleratorCopyToDeviceAsync(void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) { cudaMemcpyAsync(to,from,bytes, cudaMemcpyHostToDevice, stream);}
@@ -322,7 +335,9 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
 #define accelerator_barrier(dummy) { theGridAccelerator->wait(); }
 
 inline void *acceleratorAllocShared(size_t bytes){ return malloc_shared(bytes,*theGridAccelerator);};
+inline void *acceleratorAllocHost(size_t bytes)  { return malloc_host(bytes,*theGridAccelerator);};
 inline void *acceleratorAllocDevice(size_t bytes){ return malloc_device(bytes,*theGridAccelerator);};
+inline void acceleratorFreeHost(void *ptr){free(ptr,*theGridAccelerator);};
 inline void acceleratorFreeShared(void *ptr){free(ptr,*theGridAccelerator);};
 inline void acceleratorFreeDevice(void *ptr){free(ptr,*theGridAccelerator);};
 
@@ -441,6 +456,16 @@ void LambdaApply(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
     }								\
   }
 
+inline void *acceleratorAllocHost(size_t bytes)
+{
+  void *ptr=NULL;
+  auto err = hipMallocHost((void **)&ptr,bytes);
+  if( err != hipSuccess ) {
+    ptr = (void *) NULL;
+    fprintf(stderr," hipMallocManaged failed for %ld %s \n",bytes,hipGetErrorString(err)); fflush(stderr);
+  }
+  return ptr;
+};
 inline void *acceleratorAllocShared(size_t bytes)
 {
   void *ptr=NULL;
@@ -464,6 +489,7 @@ inline void *acceleratorAllocDevice(size_t bytes)
   return ptr;
 };
 
+inline void acceleratorFreeHost(void *ptr){ auto discard=hipFree(ptr);};
 inline void acceleratorFreeShared(void *ptr){ auto discard=hipFree(ptr);};
 inline void acceleratorFreeDevice(void *ptr){ auto discard=hipFree(ptr);};
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { auto discard=hipMemcpy(to,from,bytes, hipMemcpyHostToDevice);}
@@ -546,8 +572,10 @@ inline void acceleratorCopySynchronise(void) {};
 inline int  acceleratorIsCommunicable(void *ptr){ return 1; }
 inline void acceleratorMemSet(void *base,int value,size_t bytes) { memset(base,value,bytes);}
 #ifdef HAVE_MM_MALLOC_H
+inline void *acceleratorAllocHost(size_t bytes){return _mm_malloc(bytes,GRID_ALLOC_ALIGN);};
 inline void *acceleratorAllocShared(size_t bytes){return _mm_malloc(bytes,GRID_ALLOC_ALIGN);};
 inline void *acceleratorAllocDevice(size_t bytes){return _mm_malloc(bytes,GRID_ALLOC_ALIGN);};
+inline void acceleratorFreeHost(void *ptr){_mm_free(ptr);};
 inline void acceleratorFreeShared(void *ptr){_mm_free(ptr);};
 inline void acceleratorFreeDevice(void *ptr){_mm_free(ptr);};
 #else
diff --git a/benchmarks/Benchmark_dwf_fp32.cc b/benchmarks/Benchmark_dwf_fp32.cc
index ce4fcfab..cbe1ee23 100644
--- a/benchmarks/Benchmark_dwf_fp32.cc
+++ b/benchmarks/Benchmark_dwf_fp32.cc
@@ -52,7 +52,7 @@ int main (int argc, char ** argv)
 
   int threads = GridThread::GetThreads();
 
-  int Ls=16;
+  int Ls=8;
   for(int i=0;i<argc;i++) {
     if(std::string(argv[i]) == "-Ls"){
       std::stringstream ss(argv[i+1]); ss >> Ls;
diff --git a/systems/Aurora/benchmarks/gpu_tile.sh b/systems/Aurora/benchmarks/gpu_tile.sh
index ef64299c..a622ba3e 100755
--- a/systems/Aurora/benchmarks/gpu_tile.sh
+++ b/systems/Aurora/benchmarks/gpu_tile.sh
@@ -19,12 +19,12 @@ export ONEAPI_DEVICE_FILTER=gpu,level_zero
 
 export SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS=0
 export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
-export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0:7
+export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0:3
 export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE_FOR_D2D_COPY=1
 #export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0:2
 #export SYCL_PI_LEVEL_ZERO_USM_RESIDENT=1
 
-export MPI_BUF_NUMA=$NUMAH
+#export MPI_BUF_NUMA=$NUMAH
 
 echo "rank $PALS_RANKID ; local rank $PALS_LOCAL_RANKID ; ZE_AFFINITY_MASK=$ZE_AFFINITY_MASK ; NUMA $NUMA "