diff --git a/GPU/GPUTracking/Base/GPUParam.h b/GPU/GPUTracking/Base/GPUParam.h
index d4edadc709b8d..2d0b978241a98 100644
--- a/GPU/GPUTracking/Base/GPUParam.h
+++ b/GPU/GPUTracking/Base/GPUParam.h
@@ -38,7 +38,6 @@ namespace gpu
 struct GPUSettingsRec;
 struct GPUSettingsGTP;
 struct GPURecoStepConfiguration;
-struct GPUTPCClusterOccupancyMapBin;
 
 struct GPUParamSlice {
   float Alpha;              // slice angle
@@ -60,7 +59,7 @@ struct GPUParam_t {
 
   GPUTPCGeometry tpcGeometry;                       // TPC Geometry
   GPUTPCGMPolynomialField polynomialField;          // Polynomial approx. of magnetic field for TPC GM
-  const GPUTPCClusterOccupancyMapBin* occupancyMap; // Ptr to TPC occupancy map
+  const unsigned int* occupancyMap;                 // Ptr to TPC occupancy map
 
   GPUParamSlice SliceParam[GPUCA_NSLICES];
 
@@ -104,7 +103,7 @@ struct GPUParam : public internal::GPUParam_t<GPUSettingsRec, GPUSettingsParam>
   GPUd() float GetClusterError2(int yz, int type, float z, float angle2, float scaledMult, float scaledAvgCharge) const;
   GPUd() void GetClusterErrors2(char sector, int row, float z, float sinPhi, float DzDs, float time, float avgCharge, float& ErrY2, float& ErrZ2) const;
   GPUd() void UpdateClusterError2ByState(short clusterState, float& ErrY2, float& ErrZ2) const;
-  GPUd() float GetScaledMult(int iSlice, int iRow, float time) const;
+  GPUd() float GetScaledMult(float time) const;
 
   GPUd() void Slice2Global(int iSlice, float x, float y, float z, float* X, float* Y, float* Z) const;
   GPUd() void Global2Slice(int iSlice, float x, float y, float z, float* X, float* Y, float* Z) const;
diff --git a/GPU/GPUTracking/Base/GPUParam.inc b/GPU/GPUTracking/Base/GPUParam.inc
index 4df53e3ff6a75..52a6c76419f73 100644
--- a/GPU/GPUTracking/Base/GPUParam.inc
+++ b/GPU/GPUTracking/Base/GPUParam.inc
@@ -163,7 +163,7 @@ GPUdi() void MEM_LG(GPUParam)::GetClusterErrors2(char sector, int iRow, float z,
   float angleY2 = s2 * sec2;          // dy/dx
   float angleZ2 = DzDs * DzDs * sec2; // dz/dx
 
-  float mult = time >= 0.f ? GetScaledMult(sector, iRow, time) / tpcGeometry.Row2X(iRow) : 0.f;
+  float mult = time >= 0.f ? GetScaledMult(time) / tpcGeometry.Row2X(iRow) : 0.f;
 
   ErrY2 = GetClusterError2(0, rowType, z, angleY2, mult, avgCharge);
   ErrZ2 = GetClusterError2(1, rowType, z, angleZ2, mult, avgCharge);
@@ -191,14 +191,14 @@ GPUdi() void MEM_LG(GPUParam)::UpdateClusterError2ByState(short clusterState, fl
 }
 
 MEM_CLASS_PRE()
-GPUdi() float MEM_LG(GPUParam)::GetScaledMult(int iSlice, int iRow, float time) const
+GPUdi() float MEM_LG(GPUParam)::GetScaledMult(float time) const
 {
 #if !defined(__OPENCL__) || defined(__OPENCLCPP__)
   if (!occupancyMap) {
     return 0.f;
   }
   const unsigned int bin = CAMath::Max(0.f, time / rec.tpc.occupancyMapTimeBins);
-  return occupancyMap[bin].bin[iSlice][iRow] * rec.tpc.clusterErrorOccupancyScaler;
+  return occupancyMap[bin] * rec.tpc.clusterErrorOccupancyScaler;
 #else
   return 0.f;
 #endif
diff --git a/GPU/GPUTracking/Base/GPUReconstruction.cxx b/GPU/GPUTracking/Base/GPUReconstruction.cxx
index 896b3b2ba77d6..fcdebe0951002 100644
--- a/GPU/GPUTracking/Base/GPUReconstruction.cxx
+++ b/GPU/GPUTracking/Base/GPUReconstruction.cxx
@@ -729,6 +729,15 @@ void* GPUReconstruction::AllocateVolatileDeviceMemory(size_t size)
   return retVal;
 }
 
+void* GPUReconstruction::AllocateVolatileMemory(size_t size, bool device)
+{
+  if (device) {
+    return AllocateVolatileDeviceMemory(size);
+  }
+  mVolatileChunks.emplace_back(new char[size + GPUCA_BUFFER_ALIGNMENT]);
+  return GPUProcessor::alignPointer<GPUCA_BUFFER_ALIGNMENT>(mVolatileChunks.back().get());
+}
+
 void GPUReconstruction::ResetRegisteredMemoryPointers(GPUProcessor* proc)
 {
   for (unsigned int i = 0; i < mMemoryResources.size(); i++) {
@@ -796,6 +805,12 @@ void GPUReconstruction::ReturnVolatileDeviceMemory()
   }
 }
 
+void GPUReconstruction::ReturnVolatileMemory()
+{
+  ReturnVolatileDeviceMemory();
+  mVolatileChunks.clear();
+}
+
 void GPUReconstruction::PushNonPersistentMemory(unsigned long tag)
 {
   mNonPersistentMemoryStack.emplace_back(mHostMemoryPoolEnd, mDeviceMemoryPoolEnd, mNonPersistentIndividualAllocations.size(), tag);
diff --git a/GPU/GPUTracking/Base/GPUReconstruction.h b/GPU/GPUTracking/Base/GPUReconstruction.h
index 518f0af43aae4..dc629d86553e2 100644
--- a/GPU/GPUTracking/Base/GPUReconstruction.h
+++ b/GPU/GPUTracking/Base/GPUReconstruction.h
@@ -62,6 +62,7 @@ class GPUReconstruction
   std::shared_ptr<LibraryLoader> mMyLib = nullptr;
   std::vector<GPUMemoryResource> mMemoryResources;
   std::vector<std::unique_ptr<char[]>> mUnmanagedChunks;
+  std::vector<std::unique_ptr<char[]>> mVolatileChunks;
   std::vector<std::unique_ptr<GPUChain>> mChains;
 
  public:
@@ -205,10 +206,12 @@ class GPUReconstruction
   void AllocateRegisteredForeignMemory(short res, GPUReconstruction* rec, GPUOutputControl* control = nullptr);
   void* AllocateUnmanagedMemory(size_t size, int type);
   void* AllocateVolatileDeviceMemory(size_t size);
+  void* AllocateVolatileMemory(size_t size, bool device);
   void FreeRegisteredMemory(GPUProcessor* proc, bool freeCustom = false, bool freePermanent = false);
   void FreeRegisteredMemory(short res);
   void ClearAllocatedMemory(bool clearOutputs = true);
   void ReturnVolatileDeviceMemory();
+  void ReturnVolatileMemory();
   void PushNonPersistentMemory(unsigned long tag);
   void PopNonPersistentMemory(RecoStep step, unsigned long tag);
   void BlockStackedMemory(GPUReconstruction* rec);
diff --git a/GPU/GPUTracking/Base/GPUReconstructionCPU.cxx b/GPU/GPUTracking/Base/GPUReconstructionCPU.cxx
index 0a5064635b3a6..fca2c802a602e 100644
--- a/GPU/GPUTracking/Base/GPUReconstructionCPU.cxx
+++ b/GPU/GPUTracking/Base/GPUReconstructionCPU.cxx
@@ -375,7 +375,7 @@ unsigned int GPUReconstructionCPU::SetAndGetNestedLoopOmpFactor(bool condition,
   return mNestedLoopOmpFactor;
 }
 
-void GPUReconstructionCPU::UpdateParamOccupancyMap(const GPUTPCClusterOccupancyMapBin* mapHost, const GPUTPCClusterOccupancyMapBin* mapGPU, int stream)
+void GPUReconstructionCPU::UpdateParamOccupancyMap(const unsigned int* mapHost, const unsigned int* mapGPU, int stream)
 {
   param().occupancyMap = mapHost;
   if (IsGPU()) {
diff --git a/GPU/GPUTracking/Base/GPUReconstructionCPU.h b/GPU/GPUTracking/Base/GPUReconstructionCPU.h
index 6f72074b53f8f..387122d8358dd 100644
--- a/GPU/GPUTracking/Base/GPUReconstructionCPU.h
+++ b/GPU/GPUTracking/Base/GPUReconstructionCPU.h
@@ -136,7 +136,7 @@ class GPUReconstructionCPU : public GPUReconstructionKernels<GPUReconstructionCP
   void SetNestedLoopOmpFactor(unsigned int f) { mNestedLoopOmpFactor = f; }
   unsigned int SetAndGetNestedLoopOmpFactor(bool condition, unsigned int max);
 
-  void UpdateParamOccupancyMap(const GPUTPCClusterOccupancyMapBin* mapHost, const GPUTPCClusterOccupancyMapBin* mapGPU, int stream = -1);
+  void UpdateParamOccupancyMap(const unsigned int* mapHost, const unsigned int* mapGPU, int stream = -1);
 
  protected:
   struct GPUProcessorProcessors : public GPUProcessor {
diff --git a/GPU/GPUTracking/Definitions/GPUSettingsList.h b/GPU/GPUTracking/Definitions/GPUSettingsList.h
index 6ed9711755fbd..c5212ebca5f07 100644
--- a/GPU/GPUTracking/Definitions/GPUSettingsList.h
+++ b/GPU/GPUTracking/Definitions/GPUSettingsList.h
@@ -96,6 +96,7 @@ AddOptionRTC(maxTimeBinAboveThresholdIn1000Bin, unsigned short, 500, "", 0, "Exc
 AddOptionRTC(maxConsecTimeBinAboveThreshold, unsigned short, 200, "", 0, "Except pad from cluster finding if number of consecutive charges in a fragment is above this baseline (disable = 0)")
 AddOptionRTC(noisyPadSaturationThreshold, unsigned short, 700, "", 0, "Threshold where a timebin is considered saturated, disabling the noisy pad check for that pad")
 AddOptionRTC(occupancyMapTimeBins, unsigned short, 100, "", 0, "Number of timebins per histogram bin of occupancy map (0 = disable occupancy map)")
+AddOptionRTC(occupancyMapTimeBinsAverage, unsigned short, 2, "", 0, "Number of timebins +/- to use for the averaging")
 AddOptionRTC(trackFitCovLimit, unsigned short, 1000, "", 0, "Abort fit when y/z cov exceed the limit")
 AddOptionRTC(addErrorsCECrossing, unsigned char, 0, "", 0, "Add additional custom track errors when crossing CE, 0 = no custom errors but att 0.5 to sigma_z^2, 1 = only to cov diagonal, 2 = preserve correlations")
 AddOptionRTC(trackMergerMinPartHits, unsigned char, 10, "", 0, "Minimum hits of track part during track merging")
diff --git a/GPU/GPUTracking/Global/GPUChainTrackingSliceTracker.cxx b/GPU/GPUTracking/Global/GPUChainTrackingSliceTracker.cxx
index 8ab4cb1913b33..dbd497551f8bc 100644
--- a/GPU/GPUTracking/Global/GPUChainTrackingSliceTracker.cxx
+++ b/GPU/GPUTracking/Global/GPUChainTrackingSliceTracker.cxx
@@ -149,14 +149,16 @@ int GPUChainTracking::RunTPCTrackingSlices_internal()
     AllocateRegisteredMemory(mInputsHost->mResourceOccupancyMap, mSubOutputControls[GPUTrackingOutputs::getIndex(&GPUTrackingOutputs::tpcOccupancyMap)]);
     ReleaseEvent(mEvents->init);
     auto* ptr = doGPU ? mInputsShadow->mTPCClusterOccupancyMap : mInputsHost->mTPCClusterOccupancyMap;
+    auto* ptrTmp = (GPUTPCClusterOccupancyMapBin*)mRec->AllocateVolatileMemory(GPUTPCClusterOccupancyMapBin::getTotalSize(param()), doGPU);
     int streamOccMap = mRec->NStreams() - 1;
-    runKernel<GPUMemClean16>(GetGridAutoStep(streamOccMap, RecoStep::TPCSliceTracking), krnlRunRangeNone, {}, ptr, GPUTPCClusterOccupancyMapBin::getTotalSize(param()));
-    runKernel<GPUTPCCreateOccupancyMap, GPUTPCCreateOccupancyMap::fill>(GetGridBlk(GPUCA_NSLICES * GPUCA_ROW_COUNT, streamOccMap), krnlRunRangeNone, krnlEventNone, ptr);
-    runKernel<GPUTPCCreateOccupancyMap, GPUTPCCreateOccupancyMap::fold>(GetGridBlk(GPUCA_NSLICES * GPUCA_ROW_COUNT, streamOccMap), krnlRunRangeNone, {&mEvents->init}, ptr);
+    runKernel<GPUMemClean16>(GetGridAutoStep(streamOccMap, RecoStep::TPCSliceTracking), krnlRunRangeNone, {}, ptrTmp, GPUTPCClusterOccupancyMapBin::getTotalSize(param()));
+    runKernel<GPUTPCCreateOccupancyMap, GPUTPCCreateOccupancyMap::fill>(GetGridBlk(GPUCA_NSLICES * GPUCA_ROW_COUNT, streamOccMap), krnlRunRangeNone, krnlEventNone, ptrTmp);
+    runKernel<GPUTPCCreateOccupancyMap, GPUTPCCreateOccupancyMap::fold>(GetGridBlk(GPUTPCClusterOccupancyMapBin::getNBins(param()), streamOccMap), krnlRunRangeNone, krnlEventNone, ptrTmp, ptr);
+    mRec->ReturnVolatileMemory();
     if (doGPU) {
-      TransferMemoryResourceLinkToHost(RecoStep::TPCSliceTracking, mInputsHost->mResourceOccupancyMap);
+      TransferMemoryResourceLinkToHost(RecoStep::TPCSliceTracking, mInputsHost->mResourceOccupancyMap, streamOccMap, &mEvents->init);
     } else {
-      TransferMemoryResourceLinkToGPU(RecoStep::TPCSliceTracking, mInputsHost->mResourceOccupancyMap);
+      TransferMemoryResourceLinkToGPU(RecoStep::TPCSliceTracking, mInputsHost->mResourceOccupancyMap, streamOccMap, &mEvents->init);
     }
     mRec->UpdateParamOccupancyMap(mInputsHost->mTPCClusterOccupancyMap, mInputsShadow->mTPCClusterOccupancyMap, streamOccMap);
   }
diff --git a/GPU/GPUTracking/Global/GPUTrackingInputProvider.h b/GPU/GPUTracking/Global/GPUTrackingInputProvider.h
index 9081773ddc205..d63b3e74cf220 100644
--- a/GPU/GPUTracking/Global/GPUTrackingInputProvider.h
+++ b/GPU/GPUTracking/Global/GPUTrackingInputProvider.h
@@ -87,7 +87,7 @@ class GPUTrackingInputProvider : public GPUProcessor
   o2::tpc::ClusterNative* mPclusterNativeBuffer = nullptr;
   o2::tpc::ClusterNative* mPclusterNativeOutput = nullptr;
 
-  GPUTPCClusterOccupancyMapBin* mTPCClusterOccupancyMap = nullptr;
+  unsigned int* mTPCClusterOccupancyMap = nullptr;
 
   unsigned int* mErrorCodes = nullptr;
 
diff --git a/GPU/GPUTracking/Merger/GPUTPCGMMerger.h b/GPU/GPUTracking/Merger/GPUTPCGMMerger.h
index 82296b82a2196..5949e2ff6ad2b 100644
--- a/GPU/GPUTracking/Merger/GPUTPCGMMerger.h
+++ b/GPU/GPUTracking/Merger/GPUTPCGMMerger.h
@@ -215,7 +215,7 @@ class GPUTPCGMMerger : public GPUProcessor
   std::vector<float> StreamerUncorrectedZY(int iSlice, int iRow, const GPUTPCGMTrackParam& track, const GPUTPCGMPropagator& prop) const;
 
   void DebugStreamerUpdate(int iTrk, int ihit, float xx, float yy, float zz, const GPUTPCGMMergedTrackHit& cluster, const o2::tpc::ClusterNative& clusterNative, const GPUTPCGMTrackParam& track, const GPUTPCGMPropagator& prop, const gputpcgmmergertypes::InterpolationErrorHit& interpolation, char rejectChi2, bool refit, int retVal) const;
-  static void DebugStreamerReject(float mAlpha, int iRow, float posY, float posZ, short clusterState, char rejectChi2, const gputpcgmmergertypes::InterpolationErrorHit& inter, bool refit, int retVal, float err2Y, float err2Z, const GPUTPCGMTrackParam& track, char sector, const GPUParam& param, float time, float avgCharge);
+  static void DebugStreamerReject(float mAlpha, int iRow, float posY, float posZ, short clusterState, char rejectChi2, const gputpcgmmergertypes::InterpolationErrorHit& inter, bool refit, int retVal, float err2Y, float err2Z, const GPUTPCGMTrackParam& track, const GPUParam& param, float time, float avgCharge);
 #endif
 
   GPUdi() int SliceTrackInfoFirst(int iSlice) const { return mSliceTrackInfoIndex[iSlice]; }
diff --git a/GPU/GPUTracking/Merger/GPUTPCGMMergerDump.cxx b/GPU/GPUTracking/Merger/GPUTPCGMMergerDump.cxx
index 80dc82b1ea8a6..2a3b49417c748 100644
--- a/GPU/GPUTracking/Merger/GPUTPCGMMergerDump.cxx
+++ b/GPU/GPUTracking/Merger/GPUTPCGMMergerDump.cxx
@@ -331,11 +331,12 @@ void GPUTPCGMMerger::DebugRefitMergedTrack(const GPUTPCGMMergedTrack& track) con
 
 std::vector<unsigned short> GPUTPCGMMerger::StreamerOccupancyBin(int iSlice, int iRow, float time) const
 {
-  std::vector<unsigned short> retVal(5);
+  std::vector<unsigned short> retVal(1 + 2 * Param().rec.tpc.occupancyMapTimeBinsAverage);
 #ifdef DEBUG_STREAMER
   const int bin = CAMath::Max(0.f, time / Param().rec.tpc.occupancyMapTimeBins);
-  for (int i = 0; i < 5; i++) {
-    retVal[i] = (bin - 2 + i >= 0 && bin - 2 + i < GPUTPCClusterOccupancyMapBin::getNBins(Param())) ? Param().occupancyMap[bin - 2 + i].bin[iSlice][iRow] : 0;
+  for (int i = 0; i < 1 + 2 * Param().rec.tpc.occupancyMapTimeBinsAverage; i++) {
+    const int mybin = bin + i - Param().rec.tpc.occupancyMapTimeBinsAverage;
+    retVal[i] = (mybin >= 0 && mybin < GPUTPCClusterOccupancyMapBin::getNBins(Param())) ? Param().occupancyMap[i] : 0;
   }
 #endif
   return retVal;
@@ -375,10 +376,10 @@ void GPUTPCGMMerger::DebugStreamerUpdate(int iTrk, int ihit, float xx, float yy,
 #endif
 }
 
-void GPUTPCGMMerger::DebugStreamerReject(float mAlpha, int iRow, float posY, float posZ, short clusterState, char rejectChi2, const gputpcgmmergertypes::InterpolationErrorHit& inter, bool refit, int retVal, float err2Y, float err2Z, const GPUTPCGMTrackParam& track, char sector, const GPUParam& param, float time, float avgCharge)
+void GPUTPCGMMerger::DebugStreamerReject(float mAlpha, int iRow, float posY, float posZ, short clusterState, char rejectChi2, const gputpcgmmergertypes::InterpolationErrorHit& inter, bool refit, int retVal, float err2Y, float err2Z, const GPUTPCGMTrackParam& track, const GPUParam& param, float time, float avgCharge)
 {
 #ifdef DEBUG_STREAMER
-  float scaledMult = (time >= 0.f ? param.GetScaledMult(sector, iRow, time) / param.tpcGeometry.Row2X(iRow) : 0.f);
+  float scaledMult = (time >= 0.f ? param.GetScaledMult(time) / param.tpcGeometry.Row2X(iRow) : 0.f);
   o2::utils::DebugStreamer::instance()->getStreamer("debug_InterpolateReject", "UPDATE") << o2::utils::DebugStreamer::instance()->getUniqueTreeName("tree_InterpolateReject").data()
                                                                                          << "mAlpha=" << mAlpha
                                                                                          << "iRow=" << iRow
diff --git a/GPU/GPUTracking/Merger/GPUTPCGMPropagator.cxx b/GPU/GPUTracking/Merger/GPUTPCGMPropagator.cxx
index d7445e5fb27f3..0c3f723bc43c8 100644
--- a/GPU/GPUTracking/Merger/GPUTPCGMPropagator.cxx
+++ b/GPU/GPUTracking/Merger/GPUTPCGMPropagator.cxx
@@ -664,7 +664,7 @@ GPUd() int GPUTPCGMPropagator::Update(float posY, float posZ, int iRow, const GP
     } else {
       int retVal = InterpolateReject(param, posY, posZ, clusterState, rejectChi2, inter, err2Y, err2Z);
       GPUCA_DEBUG_STREAMER_CHECK(if (o2::utils::DebugStreamer::checkStream(o2::utils::StreamFlags::streamRejectCluster, iTrk)) {
-        GPUTPCGMMerger::DebugStreamerReject(mAlpha, iRow, posY, posZ, clusterState, rejectChi2, *inter, refit, retVal, err2Y, err2Z, *mT, sector, param, time, avgCharge);
+        GPUTPCGMMerger::DebugStreamerReject(mAlpha, iRow, posY, posZ, clusterState, rejectChi2, *inter, refit, retVal, err2Y, err2Z, *mT, param, time, avgCharge);
       });
       if (retVal) {
         return retVal;
diff --git a/GPU/GPUTracking/SliceTracker/GPUTPCCreateOccupancyMap.cxx b/GPU/GPUTracking/SliceTracker/GPUTPCCreateOccupancyMap.cxx
index 9bf1322e72ed5..84c6c65a8a7b9 100644
--- a/GPU/GPUTracking/SliceTracker/GPUTPCCreateOccupancyMap.cxx
+++ b/GPU/GPUTracking/SliceTracker/GPUTPCCreateOccupancyMap.cxx
@@ -36,33 +36,21 @@ GPUdii() void GPUTPCCreateOccupancyMap::Thread<GPUTPCCreateOccupancyMap::fill>(i
 }
 
 template <>
-GPUdii() void GPUTPCCreateOccupancyMap::Thread<GPUTPCCreateOccupancyMap::fold>(int nBlocks, int nThreads, int iBlock, int iThread, GPUsharedref() GPUSharedMemory& smem, processorType& GPUrestrict() processors, GPUTPCClusterOccupancyMapBin* GPUrestrict() map)
+GPUdii() void GPUTPCCreateOccupancyMap::Thread<GPUTPCCreateOccupancyMap::fold>(int nBlocks, int nThreads, int iBlock, int iThread, GPUsharedref() GPUSharedMemory& smem, processorType& GPUrestrict() processors, GPUTPCClusterOccupancyMapBin* GPUrestrict() map, unsigned int* GPUrestrict() output)
 {
   GPUParam& GPUrestrict() param = processors.param;
-  const int iSliceRow = iBlock * nThreads + iThread;
-  if (iSliceRow > GPUCA_ROW_COUNT * GPUCA_NSLICES) {
-    return;
-  }
-  static constexpr unsigned int FOLD_BINS_BEEFORE_AFTER = 2;
-  static constexpr unsigned int FOLD_BINS = FOLD_BINS_BEEFORE_AFTER * 2 + 1;
-  const unsigned int iSlice = iSliceRow / GPUCA_ROW_COUNT;
-  const unsigned int iRow = iSliceRow % GPUCA_ROW_COUNT;
-  const unsigned int nBins = GPUTPCClusterOccupancyMapBin::getNBins(param);
-  if (nBins < FOLD_BINS) {
+  const unsigned int bin = iBlock * nThreads + iThread;
+  if (bin >= GPUTPCClusterOccupancyMapBin::getNBins(param)) {
     return;
   }
-  unsigned short lastVal[FOLD_BINS_BEEFORE_AFTER];
-  unsigned int sum = (FOLD_BINS_BEEFORE_AFTER + 1) * map[0].bin[iSlice][iRow];
-  for (unsigned int i = 0; i < FOLD_BINS_BEEFORE_AFTER; i++) {
-    sum += map[i + 1].bin[iSlice][iRow];
-    lastVal[i] = map[0].bin[iSlice][iRow];
-  }
-  unsigned int lastValIndex = 0;
-  for (unsigned int i = 0; i < nBins; i++) {
-    unsigned short useLastVal = lastVal[lastValIndex];
-    lastVal[lastValIndex] = map[i].bin[iSlice][iRow];
-    map[i].bin[iSlice][iRow] = sum / FOLD_BINS;
-    sum += map[CAMath::Min(i + FOLD_BINS_BEEFORE_AFTER + 1, nBins - 1)].bin[iSlice][iRow] - useLastVal;
-    lastValIndex = lastValIndex < FOLD_BINS_BEEFORE_AFTER - 1 ? lastValIndex + 1 : 0;
+  int binmin = CAMath::Max<int>(0, bin - param.rec.tpc.occupancyMapTimeBinsAverage);
+  int binmax = CAMath::Min<int>(GPUTPCClusterOccupancyMapBin::getNBins(param), bin + param.rec.tpc.occupancyMapTimeBinsAverage + 1);
+  unsigned int sum = 0;
+  for (int i = binmin; i < binmax; i++) {
+    for (int iSliceRow = 0; iSliceRow < GPUCA_NSLICES * GPUCA_ROW_COUNT; iSliceRow++) {
+      sum += (&map[i].bin[0][0])[iSliceRow];
+    }
   }
+  sum /= binmax - binmin;
+  output[bin] = sum;
 }
diff --git a/GPU/GPUTracking/SliceTracker/GPUTPCCreateOccupancyMap.h b/GPU/GPUTracking/SliceTracker/GPUTPCCreateOccupancyMap.h
index 33a0d18b92a30..0cfd5dd45d9b3 100644
--- a/GPU/GPUTracking/SliceTracker/GPUTPCCreateOccupancyMap.h
+++ b/GPU/GPUTracking/SliceTracker/GPUTPCCreateOccupancyMap.h
@@ -30,8 +30,8 @@ class GPUTPCCreateOccupancyMap : public GPUKernelTemplate
            fill = 0,
            fold = 1 };
   GPUhdi() CONSTEXPR static GPUDataTypes::RecoStep GetRecoStep() { return GPUDataTypes::RecoStep::TPCSliceTracking; }
-  template <int iKernel = defaultKernel>
-  GPUd() static void Thread(int nBlocks, int nThreads, int iBlock, int iThread, GPUsharedref() GPUSharedMemory& smem, processorType& processors, GPUTPCClusterOccupancyMapBin* map);
+  template <int iKernel = defaultKernel, typename... Args>
+  GPUd() static void Thread(int nBlocks, int nThreads, int iBlock, int iThread, GPUsharedref() GPUSharedMemory& smem, processorType& processors, Args... args);
 };
 
 } // namespace GPUCA_NAMESPACE::gpu
diff --git a/GPU/GPUTracking/kernels.cmake b/GPU/GPUTracking/kernels.cmake
index ec97247ee6c79..8bdb543230ab6 100644
--- a/GPU/GPUTracking/kernels.cmake
+++ b/GPU/GPUTracking/kernels.cmake
@@ -48,7 +48,7 @@ o2_gpu_add_kernel("GPUTPCGlobalDebugSortKernels, globalTracks1"       "= TPCMERG
 o2_gpu_add_kernel("GPUTPCGlobalDebugSortKernels, globalTracks2"       "= TPCMERGER"                                           NO      single char parameter)
 o2_gpu_add_kernel("GPUTPCGlobalDebugSortKernels, borderTracks"        "= TPCMERGER"                                           NO      single char parameter)
 o2_gpu_add_kernel("GPUTPCCreateOccupancyMap, fill"                    "= TPCOCCUPANCY"                                        LB      simple GPUTPCClusterOccupancyMapBin* map)
-o2_gpu_add_kernel("GPUTPCCreateOccupancyMap, fold"                    "= TPCOCCUPANCY"                                        LB      simple GPUTPCClusterOccupancyMapBin* map)
+o2_gpu_add_kernel("GPUTPCCreateOccupancyMap, fold"                    "= TPCOCCUPANCY"                                        LB      simple GPUTPCClusterOccupancyMapBin* map "unsigned int*" output)
 o2_gpu_add_kernel("GPUTPCGMMergerTrackFit"                            "GPUTPCGMMergerGPU TPCMERGER TPCTRACKER MATLUT TPCDEDX" LB      simple int mode)
 o2_gpu_add_kernel("GPUTPCGMMergerFollowLoopers"                       "GPUTPCGMMergerGPU TPCMERGER TPCTRACKER MATLUT"         LB      simple)
 o2_gpu_add_kernel("GPUTPCGMMergerUnpackResetIds"                      "GPUTPCGMMergerGPU TPCMERGER"                           LB      simple int iSlice)