diff --git a/fpga/hls/hls_jfjoch.h b/fpga/hls/hls_jfjoch.h
index 30cce2df..2e83dfd8 100644
--- a/fpga/hls/hls_jfjoch.h
+++ b/fpga/hls/hls_jfjoch.h
@@ -70,6 +70,20 @@ struct axis_datamover_ctrl {
     ap_uint<40+64> data;
 };
 
+struct axis_completion {
+    ap_uint<128> packet_mask;
+    ap_uint<64> frame_number;
+    ap_uint<64> exptime;
+    ap_uint<64> timestamp;
+    ap_uint<64> bunchid;
+    ap_uint<32> debug;
+    ap_uint<32> data_collection_id;
+    ap_uint<32> handle;
+    ap_uint<16> packet_count;
+    ap_uint<5> module;
+    ap_uint<1> flushing;
+};
+
 void setup_datamover (hls::stream<axis_datamover_ctrl> &datamover_cmd_stream, uint64_t address, size_t bytes_to_write);
 
 void data_collection_fsm(AXI_STREAM &eth_in,
@@ -129,7 +143,7 @@ void internal_packet_generator(STREAM_512 &data_in, STREAM_512 &data_out,
 
 void save_to_hbm(STREAM_512 &data_in,
                  hls::stream<ap_uint<ADDR_STREAM_WIDTH> > &addr_in,
-                 hls::stream<ap_uint<32> > &completion_out,
+                 hls::stream<axis_completion> &completion_out,
                  hls::burst_maxi<hbm256_t> d_hbm_p0, hls::burst_maxi<hbm256_t> d_hbm_p1,
                  volatile uint64_t &packets_processed,
                  volatile ap_uint<1> &idle,
diff --git a/fpga/hls/host_writer.cpp b/fpga/hls/host_writer.cpp
index 8a923327..7ef49687 100644
--- a/fpga/hls/host_writer.cpp
+++ b/fpga/hls/host_writer.cpp
@@ -13,7 +13,7 @@ inline void write_completion(hls::stream<ap_uint<32> > &m_axis_completion,
                              const ap_uint<32> &handle,
                              const ap_uint<8> &module_number,
                              const ap_uint<64> &frame_num,
-                             const ap_uint<256> &packet_mask,
+                             const ap_uint<128> &packet_mask,
                              const ap_uint<16> &packet_count,
                              const ap_uint<32> &debug,
                              const ap_uint<64> &timestamp,
diff --git a/fpga/hls/save_to_hbm.cpp b/fpga/hls/save_to_hbm.cpp
index f4295553..f2c5460c 100644
--- a/fpga/hls/save_to_hbm.cpp
+++ b/fpga/hls/save_to_hbm.cpp
@@ -9,11 +9,11 @@
 
 #define PACKET_SIZE 8192
 
-inline void write_completion(hls::stream<ap_uint<32> > &m_axis_completion,
+inline void write_completion(hls::stream<axis_completion> &m_axis_completion,
                              const ap_uint<32> &handle,
-                             const ap_uint<8> &module_number,
+                             const ap_uint<5> &module_number,
                              const ap_uint<64> &frame_num,
-                             const ap_uint<256> &packet_mask,
+                             const ap_uint<128> &packet_mask,
                              const ap_uint<16> &packet_count,
                              const ap_uint<32> &debug,
                              const ap_uint<64> &timestamp,
@@ -22,42 +22,23 @@ inline void write_completion(hls::stream<ap_uint<32> > &m_axis_completion,
                              const ap_uint<32> &data_collection_id,
                              const ap_uint<1> &flushing) {
 #pragma HLS INLINE
-
-    ap_uint<1> all_packets_ok = packet_mask.and_reduce();
-    ap_uint<1> any_packets_received = packet_mask.or_reduce();
-    ap_uint<8> status = 0;
-    status[0] = all_packets_ok;
-    status[1] = any_packets_received;
-    status[2] = flushing;
-    ap_uint<128> tmp = (handle, packet_count, status, module_number, frame_num);
-    status[7] = tmp.xor_reduce(); // ensure completion has even parity
-
-    if (handle != HANDLE_SKIP_FRAME) {
-        m_axis_completion << handle;
-        m_axis_completion << (packet_count, status, module_number);
-        m_axis_completion << frame_num(63, 32);
-        m_axis_completion << frame_num(31,  0);
-
-        m_axis_completion << timestamp(63,32);
-        m_axis_completion << timestamp(31,0);
-        m_axis_completion << bunchid(63,32);
-        m_axis_completion << bunchid(31,0);
-
-        m_axis_completion << exptime;
-        m_axis_completion << debug;
-        m_axis_completion << 0;
-        m_axis_completion << data_collection_id;
-
-        m_axis_completion << packet_mask(127,96);
-        m_axis_completion << packet_mask( 95,64);
-        m_axis_completion << packet_mask( 63,32);
-        m_axis_completion << packet_mask( 31, 0);
-    }
+    axis_completion cmpl;
+    cmpl.handle = handle;
+    cmpl.module = module_number;
+    cmpl.frame_number = frame_num;
+    cmpl.packet_mask = packet_mask;
+    cmpl.packet_count = packet_count;
+    cmpl.debug = debug;
+    cmpl.timestamp = timestamp;
+    cmpl.bunchid = bunchid;
+    cmpl.exptime = exptime;
+    cmpl.data_collection_id = data_collection_id;
+    cmpl.flushing = flushing;
 }
 
 void save_to_hbm(STREAM_512 &data_in,
                  hls::stream<ap_uint<ADDR_STREAM_WIDTH> > &addr_in,
-                 hls::stream<ap_uint<32> > &completion_out,
+                 hls::stream<axis_completion> &completion_out,
                  hls::burst_maxi<hbm256_t> d_hbm_p0, hls::burst_maxi<hbm256_t> d_hbm_p1,
                  volatile uint64_t &packets_processed,
                  volatile ap_uint<1> &idle,
diff --git a/receiver/HLSSimulatedDevice.cpp b/receiver/HLSSimulatedDevice.cpp
index b6671a61..53e2db80 100644
--- a/receiver/HLSSimulatedDevice.cpp
+++ b/receiver/HLSSimulatedDevice.cpp
@@ -319,7 +319,7 @@ void HLSSimulatedDevice::HLSMainThread() {
     hls_cores.emplace_back([&] { writer_split(converted_2, converted_3, converted_4,
                                               addr3, addr4, addr5); });
 
-    hls::stream<ap_uint<32> > save_to_hbm_completion;
+    hls::stream<axis_completion > save_to_hbm_completion;
     ap_uint<8> save_to_hbm_err_reg;
     uint64_t save_to_hbm_packets_processed;
     ap_uint<1> save_to_hbm_idle;