diff --git a/projects/rccl/src/transport.cc b/projects/rccl/src/transport.cc
index 0f5cbf6fa6..48ae3e02d7 100644
--- a/projects/rccl/src/transport.cc
+++ b/projects/rccl/src/transport.cc
@@ -10,6 +10,7 @@
 #include "bootstrap.h"
 #define ENABLE_TIMER 0
 #include "timer.h"
+#include <cstring>
 
 struct ncclTransport* ncclTransports[NTRANSPORTS] = {
   &p2pTransport,
@@ -134,8 +135,13 @@ ncclResult_t ncclTransportP2pSetup(struct ncclComm* comm, struct ncclTopoGraph*
         struct ncclConnector* conn = comm->channels[c].peers[sendPeer].send + connIndex;
         NCCLCHECKGOTO(conn->transportComm->connect(comm, sendData++, 1, comm->rank, conn), ret, fail);
         conn->connected = 1;
-        CUDACHECKGOTO(cudaMemcpyAsync(&comm->channels[c].devPeers[sendPeer].send[connIndex], &conn->conn, sizeof(struct ncclConnInfo), cudaMemcpyHostToDevice, comm->hostStream.cudaStream), ret, fail);
-        CUDACHECKGOTO(cudaMemcpyAsync(&comm->channels[c].devPeers[sendPeer].send[connIndex], &conn->conn, sizeof(struct ncclConnInfo), cudaMemcpyHostToDevice, comm->hostStream.cudaStream), ret, fail);
+        do {
+          struct ncclConnInfo connInfo;
+          CUDACHECKGOTO(cudaMemcpyAsync(&comm->channels[c].devPeers[sendPeer].send[connIndex], &conn->conn, sizeof(struct ncclConnInfo), cudaMemcpyHostToDevice, comm->hostStream.cudaStream), ret, fail);
+          CUDACHECKGOTO(cudaMemcpyAsync(&connInfo, &comm->channels[c].devPeers[sendPeer].send[connIndex], sizeof(struct ncclConnInfo), cudaMemcpyDeviceToHost, comm->hostStream.cudaStream), ret, fail);
+          CUDACHECKGOTO(hipStreamSynchronize(comm->hostStream.cudaStream), ret, fail);
+          if (memcmp(&connInfo, &conn->conn, sizeof(struct ncclConnInfo)) == 0) break;
+        } while (true);
       }
     }
     TIME_STOP(3);
@@ -145,7 +151,13 @@ ncclResult_t ncclTransportP2pSetup(struct ncclComm* comm, struct ncclTopoGraph*
         struct ncclConnector* conn = comm->channels[c].peers[recvPeer].recv + connIndex;
         NCCLCHECKGOTO(conn->transportComm->connect(comm, recvData++, 1, comm->rank, conn), ret, fail);
         conn->connected = 1;
-        CUDACHECKGOTO(cudaMemcpyAsync(&comm->channels[c].devPeers[recvPeer].recv[connIndex], &conn->conn, sizeof(struct ncclConnInfo), cudaMemcpyHostToDevice, comm->hostStream.cudaStream), ret, fail);
+        do {
+          struct ncclConnInfo connInfo;
+          CUDACHECKGOTO(cudaMemcpyAsync(&comm->channels[c].devPeers[recvPeer].recv[connIndex], &conn->conn, sizeof(struct ncclConnInfo), cudaMemcpyHostToDevice, comm->hostStream.cudaStream), ret, fail);
+          CUDACHECKGOTO(cudaMemcpyAsync(&connInfo, &comm->channels[c].devPeers[recvPeer].recv[connIndex], sizeof(struct ncclConnInfo), cudaMemcpyDeviceToHost, comm->hostStream.cudaStream), ret, fail);
+          CUDACHECKGOTO(hipStreamSynchronize(comm->hostStream.cudaStream), ret, fail);
+          if (memcmp(&connInfo, &conn->conn, sizeof(struct ncclConnInfo)) == 0) break;
+        } while (true);
       }
     }
     TIME_STOP(4);