diff --git a/docs/changes.txt b/docs/changes.txt
index ac05e64cb..cac918b76 100644
--- a/docs/changes.txt
+++ b/docs/changes.txt
@@ -42,6 +42,7 @@
 ##
 
 - OpenCL Kernels: Add a decompressing kernel and a compressing host code in order to reduce PCIe transfer time
+- OpenCL Kernels: Improve performance preview accuracy in --benchmark, --speed-only and --progress-only mode
 - OpenCL Kernels: Remove password length restriction to 16 for Cisco-PIX and Cisco-ASA hashes
 - Terminal: Display Set Cost/Rounds During Benchmarking
 - Terminal: Show [r]esume in prompt only in pause mode, and show [p]ause in prompt only in resume mode
diff --git a/include/common.h b/include/common.h
index ac5e17dda..28ce201e5 100644
--- a/include/common.h
+++ b/include/common.h
@@ -110,7 +110,7 @@ but this is nededed for VS compiler which doesn't have inline keyword but has __
 #define PARAMCNT            64
 #define DEVICES_MAX         128
 #define EXEC_CACHE          128
-#define SPEED_CACHE         128
+#define SPEED_CACHE         4096
 #define SPEED_MAXAGE        4096
 #define EXPECTED_ITERATIONS 10000
 
diff --git a/src/hashes.c b/src/hashes.c
index b074be811..b50a25e72 100644
--- a/src/hashes.c
+++ b/src/hashes.c
@@ -330,10 +330,11 @@ void check_hash (hashcat_ctx_t *hashcat_ctx, hc_device_param_t *device_param, pl
 
 int check_cracked (hashcat_ctx_t *hashcat_ctx, hc_device_param_t *device_param, const u32 salt_pos)
 {
-  cpt_ctx_t    *cpt_ctx    = hashcat_ctx->cpt_ctx;
-  hashconfig_t *hashconfig = hashcat_ctx->hashconfig;
-  hashes_t     *hashes     = hashcat_ctx->hashes;
-  status_ctx_t *status_ctx = hashcat_ctx->status_ctx;
+  cpt_ctx_t      *cpt_ctx      = hashcat_ctx->cpt_ctx;
+  hashconfig_t   *hashconfig   = hashcat_ctx->hashconfig;
+  hashes_t       *hashes       = hashcat_ctx->hashes;
+  status_ctx_t   *status_ctx   = hashcat_ctx->status_ctx;
+  user_options_t *user_options = hashcat_ctx->user_options;
 
   salt_t *salt_buf = &hashes->salts_buf[salt_pos];
 
@@ -350,6 +351,14 @@ int check_cracked (hashcat_ctx_t *hashcat_ctx, hc_device_param_t *device_param,
     return -1;
   }
 
+  // we want the hc_clEnqueueReadBuffer to run in benchmark mode because it has an influence in performance
+  // but sometimes, when a benchmark kernel run cracks a kernel, we don't want to see that!
+
+  if (user_options->speed_only == true)
+  {
+    return 0;
+  }
+
   if (num_cracked)
   {
     plain_t *cracked = (plain_t *) hccalloc (num_cracked, sizeof (plain_t));
diff --git a/src/opencl.c b/src/opencl.c
index d9cf510c8..ca5960e7e 100644
--- a/src/opencl.c
+++ b/src/opencl.c
@@ -1318,7 +1318,7 @@ int choose_kernel (hashcat_ctx_t *hashcat_ctx, hc_device_param_t *device_param,
 
         if (user_options->speed_only == true)
         {
-          if (speed_msec > 4096) return -2; // special RC
+          if (speed_msec > 4000) return -2; // special RC
         }
       }
 
@@ -2460,27 +2460,12 @@ int run_cracker (hashcat_ctx_t *hashcat_ctx, hc_device_param_t *device_param, co
         if (CL_rc == -1) return -1;
       }
 
-      /*
-      // this writes speed cache, we dont want it
-      if (user_options->speed_only == true)
-      {
-        for (int i = 0; i < 16; i++)
-        {
-          const int rc = choose_kernel (hashcat_ctx, device_param, highest_pw_len, pws_cnt, fast_iteration, salt_pos);
-
-          if (rc == -1) return -1;
-        }
-
-        hc_timer_set (&device_param->timer_speed);
-      }
-      */
-
       const int rc = choose_kernel (hashcat_ctx, device_param, highest_pw_len, pws_cnt, fast_iteration, salt_pos);
 
       if (rc == -1) return -1;
 
       /**
-       * benchmark, part1
+       * benchmark was aborted because too long kernel runtime (slow hashes only)
        */
 
       if (user_options->speed_only == true)
@@ -2530,7 +2515,43 @@ int run_cracker (hashcat_ctx_t *hashcat_ctx, hc_device_param_t *device_param, co
        * benchmark, part2
        */
 
-      if (user_options->speed_only == true) break;
+      if (user_options->speed_only == true)
+      {
+        double total = device_param->speed_msec[0];
+
+        for (u32 speed_pos = 1; speed_pos < device_param->speed_pos; speed_pos++)
+        {
+          total += device_param->speed_msec[speed_pos];
+        }
+
+        // it's unclear if 4s is enough to turn on boost mode for all opencl device
+
+        if ((total > 4000) || (device_param->speed_pos == SPEED_CACHE - 1))
+        {
+          u32 q = device_param->speed_pos / 10; // only use the last 10% of the recording
+
+          if (q == 0) q = 1;
+
+          u64    cnt  = 0;
+          double msec = 0;
+
+          for (u32 speed_pos = device_param->speed_pos - q; speed_pos < device_param->speed_pos; speed_pos++)
+          {
+            cnt  += device_param->speed_cnt[speed_pos];
+            msec += device_param->speed_msec[speed_pos];
+          }
+
+          memset (device_param->speed_cnt,  0, SPEED_CACHE * sizeof (u64));
+          memset (device_param->speed_msec, 0, SPEED_CACHE * sizeof (double));
+
+          device_param->speed_cnt[0]  = cnt  / q;
+          device_param->speed_msec[0] = msec / q;
+
+          device_param->speed_pos = 1;
+
+          break;
+        }
+      }
 
       /**
        * result
@@ -6050,6 +6071,12 @@ void opencl_session_reset (hashcat_ctx_t *hashcat_ctx)
 
     device_param->words_off  = 0;
     device_param->words_done = 0;
+
+    #if defined (_WIN)
+    device_param->timer_speed.QuadPart = 0;
+    #else
+    device_param->timer_speed.tv_sec = 0;
+    #endif
   }
 
   opencl_ctx->kernel_power_all   = 0;