diff --git a/sdk/load-and-run/src/mgblar.cpp b/sdk/load-and-run/src/mgblar.cpp
index a30e21b1817ca7ff7afbfe0081b5c56fb478ec41..15ee3456e5acfd9ed82e02db2915cc80804c5db6 100644
--- a/sdk/load-and-run/src/mgblar.cpp
+++ b/sdk/load-and-run/src/mgblar.cpp
@@ -709,6 +709,41 @@ void run_test_st(Args &env) {
         }
     };
 
+    auto run_iters = [&](uint32_t case_idx) -> float {
+        double time_sqrsum = 0, time_sum = 0,
+               min_time = std::numeric_limits<double>::max(), max_time = 0;
+        for (int run = 0; run < env.nr_run; ++run) {
+            mgb_log_debug("load_and_run: before running iter %d", run);
+            timer.reset();
+            func->execute();
+            mgb_log_debug("load_and_run: before waiting iter %d", run);
+            auto exec_time = timer.get_msecs();
+            func->wait();
+            output_dumper.write_to_file();
+            auto cur = timer.get_msecs();
+            printf("iter %d/%d: %.3fms (exec=%.3f,device=%.3f)\n", run,
+                   env.nr_run, cur, exec_time,
+                   func->get_prev_exec_time() * 1e3);
+            time_sum += cur;
+            time_sqrsum += cur * cur;
+            fflush(stdout);
+            if (cur < min_time) {
+                min_time = cur;
+            }
+            if (cur > max_time) {
+                max_time = cur;
+            }
+        }
+        printf("=== finished test #%u: time=%.3fms avg_time=%.3fms "
+               "sd=%.3fms minmax=%.3f,%.3f\n\n",
+               case_idx, time_sum, time_sum / env.nr_run,
+               std::sqrt((time_sqrsum * env.nr_run - time_sum * time_sum) /
+                         (env.nr_run * (env.nr_run - 1))),
+               min_time, max_time);
+        return time_sum;
+
+    };
+
     if (nr_test) {
         // run testcase, generated by dump_with_testcase.py
 
@@ -742,37 +777,7 @@ void run_test_st(Args &env) {
             if (!env.nr_run) {
                 continue;
             }
-            double time_sqrsum = 0, time_sum = 0,
-                   min_time = std::numeric_limits<double>::max(), max_time = 0;
-            for (int run = 0; run < env.nr_run; ++ run) {
-                mgb_log_debug("load_and_run: before running iter %d", run);
-                timer.reset();
-                func->execute();
-                mgb_log_debug("load_and_run: before waiting iter %d", run);
-                auto exec_time = timer.get_msecs();
-                func->wait();
-                output_dumper.write_to_file();
-                auto cur = timer.get_msecs();
-                printf("iter %d/%d: %.3fms (exec=%.3f,device=%.3f)\n", run,
-                       env.nr_run, cur, exec_time,
-                       func->get_prev_exec_time() * 1e3);
-                time_sum += cur;
-                time_sqrsum += cur * cur;
-                fflush(stdout);
-                if (cur < min_time) {
-                    min_time = cur;
-                }
-                if (cur > max_time) {
-                    max_time = cur;
-                }
-            }
-            tot_time += time_sum;
-            printf("=== finished test #%u: time=%.3fms avg_time=%.3fms "
-                   "sd=%.3fms minmax=%.3f,%.3f\n\n",
-                   i, time_sum, time_sum / env.nr_run,
-                   std::sqrt((time_sqrsum * env.nr_run - time_sum * time_sum) /
-                             (env.nr_run * (env.nr_run - 1))),
-                   min_time, max_time);
+            tot_time += run_iters(i);
         }
 
         printf("=== total time: %.3fms\n", tot_time);
@@ -793,15 +798,10 @@ void run_test_st(Args &env) {
             in->copy_from(i.second);
         }
 
+        warmup();
         timer.reset();
-        func->execute();
-        auto exec_time = timer.get_msecs();
-        func->wait();
-        output_dumper.write_to_file();
-        auto cur = timer.get_msecs();
-        printf("%.3fms  %.3fms (device=%.3f)\n", cur, exec_time,
-               func->get_prev_exec_time() * 1e3);
-
+        printf("=== going to run input for %d times\n", env.nr_run);
+        run_iters(0);
     } else {
         // run speed test for a raw mgb graph
         mgb_assert(env.load_ret.tensor_map.empty(),