diff --git a/include/ck/host_utility/flush_cache.hpp b/include/ck/host_utility/flush_cache.hpp
index b1a92b77b7..a0f47959e6 100644
--- a/include/ck/host_utility/flush_cache.hpp
+++ b/include/ck/host_utility/flush_cache.hpp
@@ -351,7 +351,10 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
         }
 #else
         // return total_time / nrepeat;
-        return (total_time - 0.01 * nrepeat) / nrepeat;
+        hipDeviceProp_t deviceProps;
+        hip_check_error(hipGetDeviceProperties(&deviceProps, 0));
+        float preprocess_offset = deviceProps.multiProcessorCount==80? 0.005 : 0.01;
+        return (total_time - preprocess_offset * nrepeat) / nrepeat;
 #endif
     }
     else
diff --git a/profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp b/profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp
index 970e4a56fa..29a645e9d7 100644
--- a/profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp
+++ b/profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp
@@ -267,9 +267,9 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
 
                 float gb_per_sec = num_btype / 1.E6 / ave_time;
 
-                // std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
-                //           << " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
-                //           << kbatch_curr << std::endl;
+                std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
+                          << " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
+                          << kbatch_curr << std::endl;
 
 #if defined CK_ENABLE_FP8 || defined CK_ENABLE_INT8
                 // set softer tolerances for fp8