diff --git a/src/Makefile b/src/Makefile
index 52169bb3e1..26e653e7d6 100644
--- a/src/Makefile
+++ b/src/Makefile
@@ -1,5 +1,5 @@
 #
-# Copyright (c) 2015-2019, NVIDIA CORPORATION. All rights reserved.
+# Copyright (c) 2015-2021, NVIDIA CORPORATION. All rights reserved.
 #
 # See LICENSE.txt for license information
 #
@@ -70,7 +70,7 @@ NVLDFLAGS += $(LIBRARIES:%=-l%)
 DST_DIR := $(BUILDDIR)
 SRC_FILES := $(wildcard *.cu)
 OBJ_FILES := $(SRC_FILES:%.cu=${DST_DIR}/%.o)
-BIN_FILES_LIST := all_reduce all_gather broadcast reduce_scatter reduce alltoall
+BIN_FILES_LIST := all_reduce all_gather broadcast reduce_scatter reduce alltoall scatter sendrecv hypercube
 BIN_FILES := $(BIN_FILES_LIST:%=${DST_DIR}/%_perf)
 
 build: ${BIN_FILES}
diff --git a/src/hypercube.cu b/src/hypercube.cu
new file mode 100644
index 0000000000..142f1a6359
--- /dev/null
+++ b/src/hypercube.cu
@@ -0,0 +1,124 @@
+/*************************************************************************
+ * Copyright (c) 2016-2021, NVIDIA CORPORATION. All rights reserved.
+ *
+ * See LICENSE.txt for license information
+ ************************************************************************/
+
+#include "cuda_runtime.h"
+#include "common.h"
+
+#define ALIGN 4
+
+void print_header() {
+  PRINT("# %10s  %12s  %8s            out-of-place                       in-place          \n", "", "", "");
+  PRINT("# %10s  %12s  %8s  %7s  %6s  %6s  %5s  %7s  %6s  %6s  %5s\n", "size", "count", "type",
+        "time", "algbw", "busbw", "error", "time", "algbw", "busbw", "error");
+  PRINT("# %10s  %12s  %8s  %7s  %6s  %6s  %5s  %7s  %6s  %6s  %5s\n", "(B)", "(elements)", "",
+        "(us)", "(GB/s)", "(GB/s)", "", "(us)", "(GB/s)", "(GB/s)", "");
+}
+
+void print_line_header (size_t size, size_t count, const char *typeName, const char *opName, int root) {
+  PRINT("%12li  %12li  %8s", size, count, typeName);
+}
+
+void HyperCubeGetCollByteCount(size_t *sendcount, size_t *recvcount, size_t *paramcount, size_t *sendInplaceOffset, size_t *recvInplaceOffset, size_t count, int nranks) {
+  size_t base = (count/(ALIGN*nranks))*ALIGN;
+  *sendcount = base;
+  *recvcount = base*nranks;
+  *sendInplaceOffset = base;
+  *recvInplaceOffset = 0;
+  *paramcount = base;
+}
+
+testResult_t HyperCubeInitData(struct threadArgs* args, ncclDataType_t type, ncclRedOp_t op, int root, int rep, int in_place) {
+  size_t sendcount = args->sendBytes / wordSize(type);
+  size_t recvcount = args->expectedBytes / wordSize(type);
+  int nranks = args->nProcs*args->nThreads*args->nGpus;
+
+  for (int i=0; i<args->nGpus; i++) {
+    int gpuid = args->localRank*args->nThreads*args->nGpus + args->thread*args->nGpus + i;
+    CUDACHECK(cudaSetDevice(gpuid));
+    int rank = ((args->proc*args->nThreads + args->thread)*args->nGpus + i);
+    CUDACHECK(cudaMemset(args->recvbuffs[i], 0, args->expectedBytes));
+    void* data = in_place ? ((char*)args->recvbuffs[i])+rank*args->sendBytes : args->sendbuffs[i];
+    TESTCHECK(InitData(data, sendcount, type, rep, rank));
+    for (int j=0; j<nranks; j++) {
+      TESTCHECK(InitData(((char*)args->expected[i])+args->sendBytes*j, sendcount, type, rep, j));
+    }
+    CUDACHECK(cudaDeviceSynchronize());
+  }
+  return testSuccess;
+}
+
+void HyperCubeGetBw(size_t count, int typesize, double sec, double* algBw, double* busBw, int nranks) {
+  double baseBw = (double)(count * typesize * (nranks - 1)) / 1.0E9 / sec;
+
+  *algBw = baseBw;
+  double factor = 1;
+  *busBw = baseBw * factor;
+}
+
+testResult_t HyperCubeRunColl(void* sendbuff, void* recvbuff, size_t count, ncclDataType_t type, ncclRedOp_t op, int root, ncclComm_t comm, cudaStream_t stream) {
+  char* sbuff = (char*)sendbuff;
+  char* rbuff = (char*)recvbuff;
+  int nRanks;
+  NCCLCHECK(ncclCommCount(comm, &nRanks));
+  int rank;
+  NCCLCHECK(ncclCommUserRank(comm, &rank));
+  size_t rankSize = count * wordSize(type);
+
+  if (rbuff+rank*rankSize != sbuff) CUDACHECK(cudaMemcpyAsync(rbuff+rank*rankSize, sbuff, rankSize, cudaMemcpyDeviceToDevice, stream));
+
+  // Hypercube AllGather
+  for (int mask=1; mask<nRanks; mask<<=1) {
+    NCCLCHECK(ncclGroupStart());
+    int s = rank & ~(mask-1);
+    int r = s ^ mask;
+    NCCLCHECK(ncclSend(rbuff+s*rankSize, count*mask, type, rank^mask, comm, stream));
+    NCCLCHECK(ncclRecv(rbuff+r*rankSize, count*mask, type, rank^mask, comm, stream));
+    NCCLCHECK(ncclGroupEnd());
+  }
+  return testSuccess;
+}
+
+struct testColl hyperCubeTest = {
+  "HyperCube",
+  HyperCubeGetCollByteCount,
+  HyperCubeInitData,
+  HyperCubeGetBw,
+  HyperCubeRunColl
+};
+
+void HyperCubeGetBuffSize(size_t *sendcount, size_t *recvcount, size_t count, int nranks) {
+  size_t paramcount, sendInplaceOffset, recvInplaceOffset;
+  HyperCubeGetCollByteCount(sendcount, recvcount, &paramcount, &sendInplaceOffset, &recvInplaceOffset, count, nranks);
+}
+
+testResult_t HyperCubeRunTest(struct threadArgs* args, int root, ncclDataType_t type, const char* typeName, ncclRedOp_t op, const char* opName) {
+  args->collTest = &hyperCubeTest;
+  ncclDataType_t *run_types;
+  const char **run_typenames;
+  int type_count;
+
+  if ((int)type != -1) {
+    type_count = 1;
+    run_types = &type;
+    run_typenames = &typeName;
+  } else {
+    type_count = test_typenum;
+    run_types = test_types;
+    run_typenames = test_typenames;
+  }
+
+  for (int i=0; i<type_count; i++) {
+    TESTCHECK(TimeTest(args, run_types[i], run_typenames[i], (ncclRedOp_t)0, "", -1));
+  }
+  return testSuccess;
+}
+
+struct testEngine hyperCubeEngine = {
+  HyperCubeGetBuffSize,
+  HyperCubeRunTest
+};
+
+#pragma weak ncclTestEngine=hyperCubeEngine
diff --git a/src/scatter.cu b/src/scatter.cu
new file mode 100644
index 0000000000..93ab2e694a
--- /dev/null
+++ b/src/scatter.cu
@@ -0,0 +1,125 @@
+/*************************************************************************
+ * Copyright (c) 2016-2021, NVIDIA CORPORATION. All rights reserved.
+ *
+ * See LICENSE.txt for license information
+ ************************************************************************/
+
+#include "cuda_runtime.h"
+#include "common.h"
+
+void print_header() {
+  PRINT("# %10s  %12s  %8s  %6s            out-of-place                       in-place          \n", "", "", "", "");
+  PRINT("# %10s  %12s  %8s  %6s  %7s  %6s  %6s  %5s  %7s  %6s  %6s  %5s\n", "size", "count", "type", "root",
+        "time", "algbw", "busbw", "error", "time", "algbw", "busbw", "error");
+  PRINT("# %10s  %12s  %8s  %6s  %7s  %6s  %6s  %5s  %7s  %6s  %6s  %5s\n", "(B)", "(elements)", "", "",
+        "(us)", "(GB/s)", "(GB/s)", "", "(us)", "(GB/s)", "(GB/s)", "");
+}
+
+void print_line_header (size_t size, size_t count, const char *typeName, const char *opName, int root) {
+  PRINT("%12li  %12li  %8s  %6i", size, count, typeName, root);
+}
+
+void ScatterGetCollByteCount(size_t *sendcount, size_t *recvcount, size_t *paramcount, size_t *sendInplaceOffset, size_t *recvInplaceOffset, size_t count, int nranks) {
+  *sendcount = (count/nranks)*nranks;
+  *recvcount = count/nranks;
+  *sendInplaceOffset = 0;
+  *recvInplaceOffset = count/nranks;
+  *paramcount = count/nranks;
+}
+
+testResult_t ScatterInitData(struct threadArgs* args, ncclDataType_t type, ncclRedOp_t op, int root, int rep, int in_place) {
+  size_t sendcount = args->sendBytes / wordSize(type);
+  size_t recvcount = args->expectedBytes / wordSize(type);
+
+  for (int i=0; i<args->nGpus; i++) {
+    int gpuid = args->localRank*args->nThreads*args->nGpus + args->thread*args->nGpus + i;
+    CUDACHECK(cudaSetDevice(gpuid));
+    int rank = ((args->proc*args->nThreads + args->thread)*args->nGpus + i);
+    CUDACHECK(cudaMemset(args->recvbuffs[i], 0, args->expectedBytes));
+    void* data = in_place ? args->recvbuffs[i] : args->sendbuffs[i];
+    if (rank == root) TESTCHECK(InitData(data, sendcount, type, rep, rank));
+    TESTCHECK(InitData(args->expected[i], recvcount, type, rep+rank*recvcount, root));
+    CUDACHECK(cudaDeviceSynchronize());
+  }
+  return testSuccess;
+}
+
+void ScatterGetBw(size_t count, int typesize, double sec, double* algBw, double* busBw, int nranks) {
+  double baseBw = (double)(count * nranks * typesize) / 1.0E9 / sec;
+
+  *algBw = baseBw;
+  double factor = ((double)(nranks-1))/((double)(nranks));
+  *busBw = baseBw * factor;
+}
+
+testResult_t ScatterRunColl(void* sendbuff, void* recvbuff, size_t count, ncclDataType_t type, ncclRedOp_t op, int root, ncclComm_t comm, cudaStream_t stream) {
+  int nRanks;
+  NCCLCHECK(ncclCommCount(comm, &nRanks));
+  int rank;
+  NCCLCHECK(ncclCommUserRank(comm, &rank));
+  size_t rankOffset = count * wordSize(type);
+  if (count == 0) return testSuccess;
+
+  NCCLCHECK(ncclGroupStart());
+  if (rank == root) {
+    for (int r=0; r<nRanks; r++) {
+      NCCLCHECK(ncclSend(((char*)sendbuff)+r*rankOffset, count, type, r, comm, stream));
+    }
+  }
+  NCCLCHECK(ncclRecv(recvbuff, count, type, root, comm, stream));
+  NCCLCHECK(ncclGroupEnd());
+
+  return testSuccess;
+}
+
+struct testColl scatterTest = {
+  "Scatter",
+  ScatterGetCollByteCount,
+  ScatterInitData,
+  ScatterGetBw,
+  ScatterRunColl
+};
+
+void ScatterGetBuffSize(size_t *sendcount, size_t *recvcount, size_t count, int nranks) {
+  size_t paramcount, sendInplaceOffset, recvInplaceOffset;
+  ScatterGetCollByteCount(sendcount, recvcount, &paramcount, &sendInplaceOffset, &recvInplaceOffset, count, nranks);
+}
+
+testResult_t ScatterRunTest(struct threadArgs* args, int root, ncclDataType_t type, const char* typeName, ncclRedOp_t op, const char* opName) {
+  args->collTest = &scatterTest;
+  ncclDataType_t *run_types;
+  const char **run_typenames;
+  int type_count;
+  int begin_root, end_root;
+
+  if ((int)type != -1) {
+    type_count = 1;
+    run_types = &type;
+    run_typenames = &typeName;
+  } else {
+    type_count = test_typenum;
+    run_types = test_types;
+    run_typenames = test_typenames;
+  }
+
+  if (root != -1) {
+    begin_root = end_root = root;
+  } else {
+    begin_root = 0;
+    end_root = args->nProcs*args->nThreads*args->nGpus-1;
+  }
+
+  for (int i=0; i<type_count; i++) {
+    for (int j=begin_root; j<=end_root; j++) {
+      TESTCHECK(TimeTest(args, run_types[i], run_typenames[i], (ncclRedOp_t)0, "", j));
+    }
+  }
+  return testSuccess;
+}
+
+struct testEngine scatterEngine = {
+  ScatterGetBuffSize,
+  ScatterRunTest
+};
+
+#pragma weak ncclTestEngine=scatterEngine
diff --git a/src/sendrecv.cu b/src/sendrecv.cu
new file mode 100644
index 0000000000..8bebc48e3d
--- /dev/null
+++ b/src/sendrecv.cu
@@ -0,0 +1,127 @@
+/*************************************************************************
+ * Copyright (c) 2016-2021, NVIDIA CORPORATION. All rights reserved.
+ *
+ * See LICENSE.txt for license information
+ ************************************************************************/
+
+#include "cuda_runtime.h"
+#include "common.h"
+
+void print_header() {
+  PRINT("# %10s  %12s  %8s            out-of-place                       in-place          \n", "", "", "");
+  PRINT("# %10s  %12s  %8s  %7s  %6s  %6s  %5s  %7s  %6s  %6s  %5s\n", "size", "count", "type",
+        "time", "algbw", "busbw", "error", "time", "algbw", "busbw", "error");
+  PRINT("# %10s  %12s  %8s  %7s  %6s  %6s  %5s  %7s  %6s  %6s  %5s\n", "(B)", "(elements)", "",
+        "(us)", "(GB/s)", "(GB/s)", "", "(us)", "(GB/s)", "(GB/s)", "");
+}
+
+void print_line_header (size_t size, size_t count, const char *typeName, const char *opName, int root) {
+  PRINT("%12li  %12li  %8s", size, count, typeName);
+}
+
+void SendRecvGetCollByteCount(size_t *sendcount, size_t *recvcount, size_t *paramcount, size_t *sendInplaceOffset, size_t *recvInplaceOffset, size_t count, int nranks) {
+  *sendcount = count;
+  *recvcount = count;
+  *sendInplaceOffset = 0;
+  *recvInplaceOffset = 0;
+  *paramcount = *sendcount;
+}
+
+testResult_t SendRecvInitData(struct threadArgs* args, ncclDataType_t type, ncclRedOp_t op, int root, int rep, int in_place) {
+  size_t sendcount = args->sendBytes / wordSize(type);
+  size_t recvcount = args->expectedBytes / wordSize(type);
+  int nranks = args->nProcs*args->nThreads*args->nGpus;
+
+  for (int i=0; i<args->nGpus; i++) {
+    int gpuid = args->localRank*args->nThreads*args->nGpus + args->thread*args->nGpus + i;
+    CUDACHECK(cudaSetDevice(gpuid));
+    int rank = ((args->proc*args->nThreads + args->thread)*args->nGpus + i);
+    CUDACHECK(cudaMemset(args->recvbuffs[i], 0, args->expectedBytes));
+    void* data = in_place ? args->recvbuffs[i] : args->sendbuffs[i];
+    TESTCHECK(InitData(data, sendcount, type, rep, rank));
+    int peer = (rank-1+nranks)%nranks;
+    TESTCHECK(InitData(args->expected[i], recvcount, type, rep, peer));
+    CUDACHECK(cudaDeviceSynchronize());
+  }
+  // We don't support in-place sendrecv
+  args->reportErrors = in_place ? 0 : 1;
+  return testSuccess;
+}
+
+void SendRecvGetBw(size_t count, int typesize, double sec, double* algBw, double* busBw, int nranks) {
+  double baseBw = (double)(count * typesize) / 1.0E9 / sec;
+
+  *algBw = baseBw;
+  double factor = 1;
+  *busBw = baseBw * factor;
+}
+
+testResult_t SendRecvRunColl(void* sendbuff, void* recvbuff, size_t count, ncclDataType_t type, ncclRedOp_t op, int root, ncclComm_t comm, cudaStream_t stream) {
+  int nRanks;
+  NCCLCHECK(ncclCommCount(comm, &nRanks));
+  int rank;
+  NCCLCHECK(ncclCommUserRank(comm, &rank));
+  int recvPeer = (rank-1+nRanks) % nRanks;
+  int sendPeer = (rank+1) % nRanks;
+
+  NCCLCHECK(ncclGroupStart());
+  NCCLCHECK(ncclSend(sendbuff, count, type, sendPeer, comm, stream));
+  NCCLCHECK(ncclRecv(recvbuff, count, type, recvPeer, comm, stream));
+  NCCLCHECK(ncclGroupEnd());
+  return testSuccess;
+}
+
+struct testColl sendRecvTest = {
+  "SendRecv",
+  SendRecvGetCollByteCount,
+  SendRecvInitData,
+  SendRecvGetBw,
+  SendRecvRunColl
+};
+
+void SendRecvGetBuffSize(size_t *sendcount, size_t *recvcount, size_t count, int nranks) {
+  size_t paramcount, sendInplaceOffset, recvInplaceOffset;
+  SendRecvGetCollByteCount(sendcount, recvcount, &paramcount, &sendInplaceOffset, &recvInplaceOffset, count, nranks);
+}
+
+testResult_t SendRecvRunTest(struct threadArgs* args, int root, ncclDataType_t type, const char* typeName, ncclRedOp_t op, const char* opName) {
+  args->collTest = &sendRecvTest;
+  ncclDataType_t *run_types;
+  ncclRedOp_t *run_ops;
+  const char **run_typenames, **run_opnames;
+  int type_count, op_count;
+
+  if ((int)type != -1) {
+    type_count = 1;
+    run_types = &type;
+    run_typenames = &typeName;
+  } else {
+    type_count = test_typenum;
+    run_types = test_types;
+    run_typenames = test_typenames;
+  }
+
+  if ((int)op != -1) {
+    op_count = 1;
+    run_ops = &op;
+    run_opnames = &opName;
+  } else {
+    op_count = test_opnum;
+    run_ops = test_ops;
+    run_opnames = test_opnames;
+  }
+
+  for (int i=0; i<type_count; i++) {
+    for (int j=0; j<op_count; j++) {
+      TESTCHECK(TimeTest(args, run_types[i], run_typenames[i], run_ops[j], run_opnames[j], -1));
+    }
+  }
+  return testSuccess;
+}
+
+struct testEngine sendRecvEngine = {
+  SendRecvGetBuffSize,
+  SendRecvRunTest
+};
+
+#pragma weak ncclTestEngine=sendRecvEngine