diff --git a/executables/nr-ru.c b/executables/nr-ru.c
index 2b2ae6530d..e4f4408c4d 100644
--- a/executables/nr-ru.c
+++ b/executables/nr-ru.c
@@ -759,18 +759,22 @@ void tx_rf(RU_t *ru,int frame,int slot, uint64_t timestamp) {
     VCD_SIGNAL_DUMPER_DUMP_VARIABLE_BY_NAME( VCD_SIGNAL_DUMPER_VARIABLES_FRAME_NUMBER_TX0_RU, frame );
     VCD_SIGNAL_DUMPER_DUMP_VARIABLE_BY_NAME( VCD_SIGNAL_DUMPER_VARIABLES_TTI_NUMBER_TX0_RU, slot );
 
-    for (i=0; i<ru->nb_tx; i++)
+    for (i=0; i<ru->nb_tx; i++){
       txp[i] = (void *)&ru->common.txdata[i][(slot*fp->samples_per_slot)-sf_extension];
+    }
+
 
     VCD_SIGNAL_DUMPER_DUMP_VARIABLE_BY_NAME( VCD_SIGNAL_DUMPER_VARIABLES_TRX_TST, (timestamp-ru->openair0_cfg.tx_sample_advance)&0xffffffff );
     VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME( VCD_SIGNAL_DUMPER_FUNCTIONS_TRX_WRITE, 1 );
     // prepare tx buffer pointers
+    start_meas(&ru->tx_fhaul);
     txs = ru->rfdevice.trx_write_func(&ru->rfdevice,
                                       timestamp+ru->ts_offset-ru->openair0_cfg.tx_sample_advance-sf_extension,
                                       txp,
                                       siglen+sf_extension,
                                       ru->nb_tx,
                                       flags);
+    stop_meas(&ru->tx_fhaul);
     LOG_D(PHY,"[TXPATH] RU %d tx_rf, writing to TS %llu, frame %d, unwrapped_frame %d, subframe %d\n",ru->idx,
           (long long unsigned int)timestamp,frame,proc->frame_tx_unwrap,slot);
     VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME( VCD_SIGNAL_DUMPER_FUNCTIONS_TRX_WRITE, 0 );
@@ -873,7 +877,7 @@ static void *ru_thread_prach( void *param ) {
     #endif
           );
     }
-    VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME( VCD_SIGNAL_DUMPER_FUNCTIONS_PHY_RU_PRACH_RX, 0 ); */
+    VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME( VCD_SIGNAL_DUMPER_FUNCTIONS_PHY_RU_PRACH_RX, 0 );*/ 
     if (release_thread(&proc->mutex_prach,&proc->instance_cnt_prach,"ru_prach_thread") < 0) break;
   }
 
@@ -1227,8 +1231,8 @@ static void *ru_stats_thread(void *param) {
 
       if (ru->fh_north_asynch_in) print_meas(&ru->rx_fhaul,"rx_fhaul",NULL,NULL);
 
-      if (ru->fh_north_out) {
         print_meas(&ru->tx_fhaul,"tx_fhaul",NULL,NULL);
+      if (ru->fh_north_out) {
         print_meas(&ru->compression,"compression",NULL,NULL);
         print_meas(&ru->transport,"transport",NULL,NULL);
       }
diff --git a/openair1/SCHED_NR/nr_ru_procedures.c b/openair1/SCHED_NR/nr_ru_procedures.c
index 7fff0af177..0cedadf065 100644
--- a/openair1/SCHED_NR/nr_ru_procedures.c
+++ b/openair1/SCHED_NR/nr_ru_procedures.c
@@ -122,7 +122,7 @@ void nr_feptx_ofdm_2thread(RU_t *ru,int frame_tx,int tti_tx) {
   int i    = 0;
   int j    = 0;
   int ret  = 0;
-  int nb_antenna_ports = 8;
+  int nb_antenna_ports = 4;
   int ofdm_mask_full   = (1<<(ru->nb_tx*2))-1;
 
   start_meas(&ru->ofdm_total_stats);
@@ -278,6 +278,7 @@ void nr_init_feptx_thread(RU_t *ru) {
     pthread_cond_init( &feptx[i].cond_feptx, NULL);
 
     threadCreate(&feptx[i].pthread_feptx, nr_feptx_thread, (void*)&feptx[i], "feptx", -1, OAI_PRIORITY_RT);
+    LOG_I(PHY,"init feptx thread %d\n", i);
   }
 
 }
diff --git a/targets/ARCH/USRP/USERSPACE/LIB/usrp_lib.cpp b/targets/ARCH/USRP/USERSPACE/LIB/usrp_lib.cpp
index 76ebd556ca..0175df9b21 100644
--- a/targets/ARCH/USRP/USERSPACE/LIB/usrp_lib.cpp
+++ b/targets/ARCH/USRP/USERSPACE/LIB/usrp_lib.cpp
@@ -455,14 +455,14 @@ static int trx_usrp_write(openair0_device *device, openair0_timestamp timestamp,
 #if defined(__x86_64) || defined(__i386__)
   #ifdef __AVX2__
       nsamps2 = (nsamps+7)>>3;
-      __m256i buff_tx[2][nsamps2];
+      __m256i buff_tx[8][nsamps2];
   #else
     nsamps2 = (nsamps+3)>>2;
-    __m128i buff_tx[2][nsamps2];
+    __m128i buff_tx[8][nsamps2];
   #endif
 #elif defined(__arm__)
     nsamps2 = (nsamps+3)>>2;
-    int16x8_t buff_tx[2][nsamps2];
+    int16x8_t buff_tx[8][nsamps2];
 #else
     #error Unsupported CPU architecture, USRP device cannot be built
 #endif
@@ -565,14 +565,14 @@ static int trx_usrp_read(openair0_device *device, openair0_timestamp *ptimestamp
 #if defined(__x86_64) || defined(__i386__)
 #ifdef __AVX2__
     nsamps2 = (nsamps+7)>>3;
-    __m256i buff_tmp[2][nsamps2];
+    __m256i buff_tmp[8][nsamps2];
 #else
     nsamps2 = (nsamps+3)>>2;
-    __m128i buff_tmp[2][nsamps2];
+    __m128i buff_tmp[8][nsamps2];
 #endif
 #elif defined(__arm__)
     nsamps2 = (nsamps+3)>>2;
-    int16x8_t buff_tmp[2][nsamps2];
+    int16x8_t buff_tmp[8][nsamps2];
 #endif
 
     if (device->type == USRP_B200_DEV) {
diff --git a/targets/PROJECTS/GENERIC-LTE-EPC/CONF/gnb.band78.tm1.106PRB.usrpn300.conf b/targets/PROJECTS/GENERIC-LTE-EPC/CONF/gnb.band78.tm1.106PRB.usrpn300.conf
index ad88a0b89f..f3a0362a44 100644
--- a/targets/PROJECTS/GENERIC-LTE-EPC/CONF/gnb.band78.tm1.106PRB.usrpn300.conf
+++ b/targets/PROJECTS/GENERIC-LTE-EPC/CONF/gnb.band78.tm1.106PRB.usrpn300.conf
@@ -246,15 +246,16 @@ L1s = (
 RUs = (
     {		  
        local_rf       = "yes"
-         nb_tx          = 1
-         nb_rx          = 1
-         att_tx         = 0
+         nb_tx          = 4; 
+         nb_rx          = 1; 
+         att_tx         = 0; 
          att_rx         = 0;
          bands          = [7];
          max_pdschReferenceSignalPower = -27;
          max_rxgain                    = 114;
          eNB_instances  = [0];
-	 sdr_addrs = "addr=192.168.10.2,second_addr=192.168.20.2,mgmt_addr=192.168.10.2,clock_source=external,time_source=external";
+
+	 sdr_addrs = "addr=192.168.50.2,second_addr=192.168.30.2,clock_source=external,time_source=external";
 
     }
 );