diff --git a/Jobs/JobAdvectionLimCart b/Jobs/JobAdvectionLimCart
index be9b508..d2d2c24 100755
--- a/Jobs/JobAdvectionLimCart
+++ b/Jobs/JobAdvectionLimCart
@@ -1,4 +1,4 @@
-mpirun -n 6 julia --project Examples/testAdvectionCart.jl \
+mpirun -n 2 julia --project Examples/testAdvectionCart.jl \
   --Problem="LimAdvectionCart" \
   --Device="CPU" \
   --FloatTypeBackend="Float64" \
diff --git a/Jobs/JobAdvectionLimCart4 b/Jobs/JobAdvectionLimCart4
new file mode 100755
index 0000000..6df9548
--- /dev/null
+++ b/Jobs/JobAdvectionLimCart4
@@ -0,0 +1,31 @@
+export UCX_ERROR_SIGNALS=""
+srun -n 4 gpu_wrapper.sh -n 4 -e "julia --project Examples/testAdvectionCart.jl \
+  --Problem="LimAdvectionCart" \
+  --Device="GPU" \
+  --GPUType="CUDA" \
+  --NumberThreadGPU=1024 \
+  --FloatTypeBackend="Float64" \
+  --NumV=5 \
+  --NumTr=1 \
+  --HorLimit=true \
+  --Upwind=true \
+  --vtkFileName="LimAdvectionCart" \
+  --SimTime=0.0  \
+  --PrintTime=0.1 \
+  --dtau=0.0025 \
+  --IntMethod="SSPRungeKutta" \
+  --Table="SSP32" \
+  --Lx=0 \
+  --Ly=0 \
+  --H=0 \
+  --x0=0 \
+  --y0=0 \
+  --nx=40 \
+  --ny=40 \
+  --nz=40 \
+  --OrdPoly=4 \
+  --BoundaryWE="Period" \
+  --BoundarySN="Period" \
+  --BoundaryBT="" \
+  --HyperVisc=true \
+  --HyperDDiv=1.e-4"
diff --git a/Jobs/JobNHBaroWaveDrySphere4 b/Jobs/JobNHBaroWaveDrySphere4
new file mode 100755
index 0000000..36a01e7
--- /dev/null
+++ b/Jobs/JobNHBaroWaveDrySphere4
@@ -0,0 +1,44 @@
+export UCX_ERROR_SIGNALS=""
+srun -n 4 gpu_wrapper.sh -n 4 -e "julia --project Examples/testNHSphere.jl \
+  --Problem="BaroWaveDrySphere" \
+  --Device="GPU" \
+  --GPUType="CUDA" \
+  --FloatTypeBackend="Float32" \
+  --NumberThreadGPU=1024 \
+  --NumV=5 \
+  --NumTr=0 \
+  --ProfpBGrd="" \
+  --ProfRhoBGrd="" \
+  --Source=false \
+  --Forcing=false \
+  --Curl=false \
+  --ModelType="VectorInvariant" \
+  --Coriolis=true \
+  --VerticalDiffusion=false \
+  --Upwind=true \
+  --HorLimit=false \
+  --Buoyancy=true \
+  --Decomp="EqualArea" \
+  --SimDays=10 \
+  --SimSeconds=0 \
+  --PrintSeconds=0 \
+  --PrintMinutes=0 \
+  --PrintHours=0 \
+  --PrintDays=0 \
+  --StartAverageDays=100 \
+  --Flat=true \
+  --dtau=150 \
+  --IntMethod="Rosenbrock" \
+  --Table="SSP-Knoth" \
+  --TopoS="" \
+  --Stretch=true \
+  --StretchType="Exp" \
+  --GridType="CubedSphere" \
+  --nz=64 \
+  --nPanel=30 \
+  --H=30000.0 \
+  --OrdPoly=3 \
+  --HyperVisc=true \
+  --HyperDCurl=5.e14 \
+  --HyperDGrad=5.e14 \
+  --HyperDDiv=5.e14"
diff --git a/src/Examples/parameters.jl b/src/Examples/parameters.jl
index 9a8dcc6..7197078 100644
--- a/src/Examples/parameters.jl
+++ b/src/Examples/parameters.jl
@@ -244,28 +244,26 @@ Base.@kwdef struct LimAdvectionCart{FT}
   u0::FT = π / 2              # angular velocity
   r0::FT = (xmax - xmin) / 6  # bells radius
   end_time::FT = 2π           # simulation period in seconds
-  centers1xC = xmin + (xmax - xmin) / 4
-  centers1yC = ymin + (ymax - ymin) / 2
-  centers1zC = zmin + (zmax - zmin) / 2
-  centers2xC = xmin + 3 * (xmax - xmin) / 4
-  centers2yC = ymin + (ymax - ymin) / 2
-  centers2zC = zmin + (zmax - zmin) / 2
+  centers1xC::FT = xmin + (xmax - xmin) / 4
+  centers1yC::FT = ymin + (ymax - ymin) / 2
+  centers1zC::FT = zmin + (zmax - zmin) / 2
+  centers2xC::FT = xmin + 3 * (xmax - xmin) / 4
+  centers2yC::FT = ymin + (ymax - ymin) / 2
+  centers2zC::FT = zmin + (zmax - zmin) / 2
 end
 
-Base.@kwdef struct ParamAdvectionCubeRotCart{FT}
+Base.@kwdef struct ParamAdvectionCubeRotCart
   StreamFun::Bool = false
-  uMax::FT = 1.0
-  vMax::FT = 1.0
-  xC::FT = 500.0
-  zC::FT = 500.0
-  x1::FT = 299.0
-  x2::FT = 501.0
-  y1::FT = 299.0
-  y2::FT = 501.0
-  z1::FT = 299.0
-  z2::FT = 501.0
-  EndTime::FT = 1000.0
-  H::FT = 1000.0
+  uMax::Float64 = 1.0
+  vMax::Float64 = 0.0
+  xC::Float64 = 500.0
+  zC::Float64 = 500.0
+  x1::Float64 = 299.0
+  x2::Float64 = 501.0
+  z1::Float64 = 299.0
+  z2::Float64 = 501.0
+  EndTime::Float64 = 1000.0
+  H::Float64 = 1000.0
 end
 
 Base.@kwdef struct ParamAdvectionCart
@@ -315,10 +313,10 @@ function Parameters(FT,Problem::String)
     Param = ParamAdvectionCubeCart()
   elseif Problem == "AdvectionCubeRotCart"
     @show Problem
-    Param = ParamAdvectionCubeRotCart{FT}()
+    Param = ParamAdvectionCubeRotCart()
   elseif Problem == "LimAdvectionCart"
     @show Problem
-    Param = LimAdvectionCart{FT}()
+    Param = LimAdvectionCart{FT}()  
   elseif Problem == "WarmBubble2DXCart"
     @show Problem
     Param = ParamWarmBubble2DXCart()
diff --git a/src/GPU/FcnGPU.jl b/src/GPU/FcnGPU.jl
index 1ff33e5..a5164a4 100644
--- a/src/GPU/FcnGPU.jl
+++ b/src/GPU/FcnGPU.jl
@@ -59,8 +59,8 @@ function FcnAdvectionGPU!(F,U,time,FE,Metric,Phys,Cache,Exchange,Global,Param,Pr
   if Global.Model.HorLimit
     @views KLimitKernel!(DoF,qMin,qMax,U[:,:,NumV+1:NumV+NumTr],Rho,Glob,ndrange=ndrangeL)
     KernelAbstractions.synchronize(backend)
-    Parallels.ExchangeDataFSend(qMin,qMax,Exchange)
-    Parallels.ExchangeDataFRecv!(qMin,qMax,Exchange)  
+    Parallels.ExchangeDataFSendGPU(qMin,qMax,Exchange)
+    Parallels.ExchangeDataFRecvGPU!(qMin,qMax,Exchange)  
   end
 
 
diff --git a/src/Parallels/Exchange.jl b/src/Parallels/Exchange.jl
index 4121d12..57aa03a 100644
--- a/src/Parallels/Exchange.jl
+++ b/src/Parallels/Exchange.jl
@@ -1,10 +1,11 @@
 mutable struct ExchangeStruct{FT<:AbstractFloat,
                               IT1<:AbstractArray,
-                              AT3<:AbstractArray}
+                              AT3<:AbstractArray,
+                              AT4<:AbstractArray}
   IndSendBuffer::Dict{Int,IT1}
-  IndSendBufferF::Dict{Int,Array{Int,1}}
+  IndSendBufferF::Dict{Int,IT1}
   IndRecvBuffer::Dict{Int,IT1}
-  IndRecvBufferF::Dict{Int,Array{Int,1}}
+  IndRecvBufferF::Dict{Int,IT1}
   NeiProc::Array{Int, 1}
   Proc::Int
   ProcNumber::Int
@@ -13,13 +14,13 @@ mutable struct ExchangeStruct{FT<:AbstractFloat,
   SendBuffer::Dict
   #SendBuffer3::Dict{Int,Array{FT, 3}}
   SendBuffer3::Dict{Int,AT3}
-  SendBufferF::Dict{Int,Array{FT, 4}}
+  SendBufferF::Dict{Int,AT4}
   InitRecvBuffer::Bool
   InitRecvBufferF::Bool
   RecvBuffer::Dict
 # RecvBuffer3::Dict{Int,Array{FT, 3}}
   RecvBuffer3::Dict{Int,AT3}
-  RecvBufferF::Dict{Int,Array{FT, 4}}
+  RecvBufferF::Dict{Int,AT4}
   sreq::MPI.UnsafeMultiRequest
   rreq::MPI.UnsafeMultiRequest
 end
@@ -45,10 +46,12 @@ function ExchangeStruct{FT}(backend) where FT<:AbstractFloat
   sreq = MPI.UnsafeMultiRequest(0)
   rreq = MPI.UnsafeMultiRequest(0)
   AT3 = KernelAbstractions.zeros(backend,FT,0,0,0)
+  AT4 = KernelAbstractions.zeros(backend,FT,0,0,0,0)
   IT1 = KernelAbstractions.zeros(backend,Int,0)
   return ExchangeStruct{FT,
                         typeof(IT1),
-                        typeof(AT3)}(
+                        typeof(AT3),
+                        typeof(AT4)}(
     IndSendBuffer,
     IndSendBufferF,
     IndRecvBuffer,
@@ -358,6 +361,7 @@ function ExchangeStruct{FT}(backend,SubGrid,OrdPoly,CellToProc,Proc,ProcNumber,H
 
   # Copy from CPU to device
   AT3 = KernelAbstractions.zeros(backend,FT,0,0,0)
+  AT4 = KernelAbstractions.zeros(backend,FT,0,0,0,0)
   IT1 = KernelAbstractions.zeros(backend,Int,0)
 
   SendBuffer = Dict()
@@ -374,7 +378,8 @@ function ExchangeStruct{FT}(backend,SubGrid,OrdPoly,CellToProc,Proc,ProcNumber,H
 
   return ExchangeStruct{FT,
                         typeof(IT1),
-                        typeof(AT3)}(
+                        typeof(AT3),
+                        typeof(AT4)}(
     SendBuffer,
     IndSendBufferF,
     RecvBuffer,
@@ -675,6 +680,75 @@ function ExchangeDataFSend(cFMin,cFMax,Exchange)
   end
 end
 
+function ExchangeDataFSendGPU(cFMin,cFMax,Exchange)
+  backend = get_backend(cFMin)
+  FT = eltype(cFMin)
+  IndSendBufferF = Exchange.IndSendBufferF
+  IndRecvBufferF = Exchange.IndRecvBufferF
+  NeiProc = Exchange.NeiProc
+  Proc = Exchange.Proc
+  ProcNumber = Exchange.ProcNumber
+  nz = size(cFMin,1)
+  nT = size(cFMin,3)
+  if Exchange.InitRecvBufferF
+    @inbounds for iP in NeiProc
+      Exchange.RecvBufferF[iP] = zeros(nz,length(IndRecvBufferF[iP]),2,nT)
+      Exchange.SendBufferF[iP] = zeros(nz,length(IndSendBufferF[iP]),2,nT)
+    end
+    RecvBufferF = Exchange.RecvBufferF
+    SendBufferF = Exchange.SendBufferF
+    Exchange.InitRecvBufferF = false
+    Exchange.InitSendBufferF = false
+    rreq = Exchange.rreq
+    sreq = Exchange.sreq
+  else
+    RecvBufferF = Exchange.RecvBufferF
+    SendBufferF = Exchange.SendBufferF
+    rreq = Exchange.rreq
+    sreq = Exchange.sreq
+  end
+
+  group = (nz,5,1)
+  KExchangeDataFSendKernel! = ExchangeDataFSendKernel!(backend,group)
+  @inbounds for iP in NeiProc
+    ndrange = (nz,length(IndSendBufferF[iP]),nT)
+    KExchangeDataFSendKernel!(cFMin,cFMax,SendBufferF[iP],IndSendBufferF[iP],ndrange=ndrange)
+  end
+#  @inbounds for iP in NeiProc
+#    i = 0
+#    @views @inbounds for Ind in IndSendBufferF[iP]
+#      i += 1
+#      @views @. SendBufferF[iP][:,i,1,:] = cFMin[:,Ind,:]
+#      @views @. SendBufferF[iP][:,i,2,:] = cFMax[:,Ind,:]
+#    end
+#  end
+  i = 0
+  @inbounds for iP in NeiProc
+    tag = Proc + ProcNumber*iP
+    i += 1
+    @views MPI.Irecv!(RecvBufferF[iP], iP - 1, tag, MPI.COMM_WORLD, rreq[i])
+  end
+  i = 0
+  @inbounds for iP in NeiProc
+    tag = iP + ProcNumber*Proc
+    i += 1
+    @views MPI.Isend(SendBufferF[iP], iP - 1, tag, MPI.COMM_WORLD, sreq[i])
+  end
+end
+
+@kernel function ExchangeDataFSendKernel!(cFMin,cFMax,SendBufferF,IndSendBufferF)
+
+  Iz,I,IT = @index(Global, NTuple)
+  NumInd = @uniform @ndrange()[2]
+  NT = @uniform @ndrange()[3]
+
+  if I <= NumInd && IT <= NT
+    @inbounds Ind = IndSendBufferF[I]    
+    @inbounds SendBufferF[Iz,I,1,IT] = cFMin[Iz,Ind,IT]
+    @inbounds SendBufferF[Iz,I,2,IT] = cFMax[Iz,Ind,IT]
+  end  
+end
+
 function ExchangeDataFRecv!(cFMin,cFMax,Exchange)
 
   IndRecvBufferF = Exchange.IndRecvBufferF
@@ -697,6 +771,52 @@ function ExchangeDataFRecv!(cFMin,cFMax,Exchange)
   end
 end
 
+function ExchangeDataFRecvGPU!(cFMin,cFMax,Exchange)
+  backend = get_backend(cFMin)
+  FT = eltype(cFMin)
+
+  IndRecvBufferF = Exchange.IndRecvBufferF
+  NeiProc = Exchange.NeiProc
+  RecvBufferF = Exchange.RecvBufferF
+  rreq = Exchange.rreq
+  sreq = Exchange.sreq
+
+  stats = MPI.Waitall(rreq)
+  stats = MPI.Waitall(sreq)
+  MPI.Barrier(MPI.COMM_WORLD)
+  Nz = size(cFMin,1)
+  nT = size(cFMin,4)
+  group = (Nz,5,1)
+  KExchangeDataFRecvKernel! = ExchangeDataFRecvKernel!(backend,group)
+
+  #Receive
+  @inbounds for iP in NeiProc
+    ndrange = (Nz,length(IndRecvBufferF[iP]),nT)
+    KExchangeDataFRecvKernel!(cFMin,cFMax,RecvBufferF[iP],IndRecvBufferF[iP],ndrange=ndrange)
+  end
+# #Receive
+# @inbounds for iP in NeiProc
+#   i = 0
+#   @inbounds for Ind in IndRecvBufferF[iP]
+#     i += 1
+#     @views @. cFMin[:,Ind,:] = RecvBufferF[iP][:,i,1,:]
+#     @views @. cFMax[:,Ind,:] = RecvBufferF[iP][:,i,2,:]
+#   end
+# end
+end
+
+@kernel function ExchangeDataFRecvKernel!(cFMin,cFMax,RecvBufferF,IndRecvBufferF)
+
+  Iz,I,IT = @index(Global, NTuple)
+  NumInd = @uniform @ndrange()[2]
+  NT = @uniform @ndrange()[3]
+  if I <= NumInd && IT <= NT
+    @inbounds Ind = IndRecvBufferF[I]    
+    cFMin[Iz,Ind,IT] = RecvBufferF[Iz,I,1,IT]
+    cFMax[Iz,Ind,IT] = RecvBufferF[Iz,I,2,IT]
+  end  
+end  
+
 function InitExchangeData3D(backend,FT,nz,nT,Exchange) 
   IndSendBuffer = Exchange.IndSendBuffer
   IndRecvBuffer = Exchange.IndRecvBuffer