diff --git a/raven_spn/raven_spn b/raven_spn/raven_spn
index eedd7a3..815c11e 100755
Binary files a/raven_spn/raven_spn and b/raven_spn/raven_spn differ
diff --git a/raven_spn/src/raven_spn.cpp b/raven_spn/src/raven_spn.cpp
index 8d93a03..9bbe075 100644
--- a/raven_spn/src/raven_spn.cpp
+++ b/raven_spn/src/raven_spn.cpp
@@ -16,17 +16,40 @@ bool double_equals(double a, double b, double epsilon = 0.001)
     return std::abs(a - b) < epsilon;
 }
 
-void run_xspn() {
+void run_xspn(int in_addr, int out_addr) {
 	spn::mode_reg() = 0;
-	spn::input_length_reg() = 100; 		// each sample consists of 5 uint8 values
-	spn::input_addr_reg() = 0x800C0000;
-	spn::output_addr_reg() = 0x800A0000;
-	spn::num_of_in_beats_reg() = 8;		// Number of AXI4 burst beats needed to load all input data
-	spn::num_of_out_beats_reg() = 13;	// Number of AXI4 burst beats needed to store all result data
+	spn::input_length_reg() = 500; 		// each sample consists of 5 uint8 values
+	spn::input_addr_reg() = in_addr;
+	spn::output_addr_reg() = out_addr;
+	spn::num_of_in_beats_reg() = 40;		// Number of AXI4 burst beats needed to load all input data
+	spn::num_of_out_beats_reg() = 64;	// Number of AXI4 burst beats needed to store all result data
 	spn::start_reg() = 1;
+}
 
-	wait_for_interrupt();
-	spn::interrupt_reg() = 1;
+void load_input_data(int addr, const int k=0) {
+    int * mem_base = (int *) addr;
+	int offset = k * 5;
+
+	for (size_t i = 0, j = 0; j < 625; i += 4, j++) { // load 5000 samples
+		*(mem_base + j) = input_data.at(offset + i);
+		*(mem_base + j) |= (input_data.at(offset + i + 1) << 8) & 0x0000FF00;
+		*(mem_base + j) |= (input_data.at(offset + i + 2) << 16) & 0x00FF0000;
+		*(mem_base + j) |= (input_data.at(offset + i + 3) << 24) & 0xFF000000;
+	}
+}
+
+bool check_results(int addr, int k) {
+	bool result = 0;
+	int  step = 500;
+	double *res_base = (double*) (addr);
+	for (int i = 0; i < step; i++) {
+		if (!double_equals(res_base[i] * ln2, ref_data.at(k + i))) {
+			printf("XSPN ref %d comparison FAILED\n", k + i);
+			result = 1;
+		}
+	}
+    printf("Compared samples %d - %d with the reference\n", k, k+step);
+	return result;
 }
 
 /*! \brief main function
@@ -34,37 +57,33 @@ void run_xspn() {
  */
 int main() {
     platform_init();
+    int ret_val = 0;
+
     spn::mode_reg() = 1;
     spn::start_reg() = 1;
 	wait_for_interrupt();
     spn::interrupt_reg() = 1;
     printf("READOUT HW:0x%x\n", spn::readout_reg());
 
-	for (int k = 0; k < 10000; k+=100) {
-        // write input samples into the memory
-        int * mem_base = (int *) 0x800C0000;
-        int offset = k * 5;
+	int in_addr  = 0x800C0000;
+	int out_addr = 0x800A0000;
+    load_input_data(in_addr);
 
-        for(size_t i = 0, j = 0; j < 125; i += 4, j++) {
-        	*(mem_base+j) = input_data.at(offset+i);
-        	*(mem_base+j) |= (input_data.at(offset+i+1) << 8)  & 0x0000FF00;
-        	*(mem_base+j) |= (input_data.at(offset+i+2) << 16) & 0x00FF0000;
-        	*(mem_base+j) |= (input_data.at(offset+i+3) << 24) & 0xFF000000;
-        }
+    int step = 500; // number of samples to be process at once
+    for (int k = 0; k < 10000; k+=step) {
 
-        printf("XSPN process samples %d - %d\n", k, k+100);
-    	run_xspn();
+        run_xspn(in_addr, out_addr);
+        printf("XSPN processes samples %d - %d\n", k, k+step);
 
-    	// read calculation results from the memory
-        double * res_base = (double*) 0x800A0000;
+        // pre-load data for the next run while HW accelerator is running
+        if(k<(10000-step))
+            load_input_data(in_addr, k+step);
 
-        for (int i = 0; i < 100; i++) {
-            if (!double_equals(res_base[i] * ln2, ref_data.at(k+i))) {
-                printf("XSPN ref %d comparison FAILED\n", k+i);
-            }
-    	}
+        wait_for_interrupt();
+    	spn::interrupt_reg() = 1;
 
+		ret_val = check_results(out_addr, k);
 	}
 
-	return 0;
+	return ret_val;
 }