diff --git a/include/backend.h b/include/backend.h
index 074a0cd9f..e2556d7b2 100644
--- a/include/backend.h
+++ b/include/backend.h
@@ -68,7 +68,7 @@ int hc_cuMemcpyHtoD              (hashcat_ctx_t *hashcat_ctx, CUdeviceptr dstDev
 int hc_cuMemFree                 (hashcat_ctx_t *hashcat_ctx, CUdeviceptr dptr);
 int hc_cuModuleGetFunction       (hashcat_ctx_t *hashcat_ctx, CUfunction *hfunc, CUmodule hmod, const char *name);
 int hc_cuModuleLoadDataEx        (hashcat_ctx_t *hashcat_ctx, CUmodule *module, const void *image, unsigned int numOptions, CUjit_option *options, void **optionValues);
-int hc_cuModuleLoadDataExLog     (hashcat_ctx_t *hashcat_ctx, CUmodule *module, const void *image);
+int hc_cuModuleLoadDataExLog     (hashcat_ctx_t *hashcat_ctx, CUmodule *module, const void *image, const u64 threads_per_block);
 int hc_cuModuleUnload            (hashcat_ctx_t *hashcat_ctx, CUmodule hmod);
 int hc_cuStreamCreate            (hashcat_ctx_t *hashcat_ctx, CUstream *phStream, unsigned int Flags);
 int hc_cuStreamDestroy           (hashcat_ctx_t *hashcat_ctx, CUstream hStream);
diff --git a/src/backend.c b/src/backend.c
index 1d52163a9..901e0b810 100644
--- a/src/backend.c
+++ b/src/backend.c
@@ -1297,32 +1297,45 @@ int hc_cuModuleLoadDataEx (hashcat_ctx_t *hashcat_ctx, CUmodule *module, const v
   return 0;
 }
 
-int hc_cuModuleLoadDataExLog (hashcat_ctx_t *hashcat_ctx, CUmodule *module, const void *image)
+int hc_cuModuleLoadDataExLog (hashcat_ctx_t *hashcat_ctx, CUmodule *module, const void *image, const u64 threads_per_block)
 {
   #define LOG_SIZE 8192
 
   char *info_log  = hcmalloc (LOG_SIZE);
   char *error_log = hcmalloc (LOG_SIZE);
 
-  CUjit_option opts[6];
+  CUjit_option opts[7];
+  void *vals[7];
 
   opts[0] = CU_JIT_TARGET_FROM_CUCONTEXT;
-  opts[1] = CU_JIT_LOG_VERBOSE;
-  opts[2] = CU_JIT_INFO_LOG_BUFFER;
-  opts[3] = CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES;
-  opts[4] = CU_JIT_ERROR_LOG_BUFFER;
-  opts[5] = CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES;
-
-  void *vals[6];
-
   vals[0] = (void *) 0;
+
+  opts[1] = CU_JIT_LOG_VERBOSE;
   vals[1] = (void *) 1;
+
+  opts[2] = CU_JIT_INFO_LOG_BUFFER;
   vals[2] = (void *) info_log;
+
+  opts[3] = CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES;
   vals[3] = (void *) LOG_SIZE;
+
+  opts[4] = CU_JIT_ERROR_LOG_BUFFER;
   vals[4] = (void *) error_log;
+
+  opts[5] = CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES;
   vals[5] = (void *) LOG_SIZE;
 
-  const int rc_cuModuleLoadDataEx = hc_cuModuleLoadDataEx (hashcat_ctx, module, image, 6, opts, vals);
+  int opts_cnt = 6;
+
+  if (threads_per_block)
+  {
+    opts[6] = CU_JIT_THREADS_PER_BLOCK;
+    vals[6] = (void *) threads_per_block;
+
+    opts_cnt++;
+  }
+
+  const int rc_cuModuleLoadDataEx = hc_cuModuleLoadDataEx (hashcat_ctx, module, image, opts_cnt, opts, vals);
 
   #if defined (DEBUG)
   printf ("cuModuleLoadDataEx() Info Log (%d):\n%s\n\n",  (int) strlen (info_log),  info_log);
@@ -7765,9 +7778,7 @@ int backend_session_begin (hashcat_ctx_t *hashcat_ctx)
 
           if (rc_nvrtcDestroyProgram == -1) return -1;
 
-          // tbd: check for some useful options
-
-          const int rc_cuModuleLoadDataEx = hc_cuModuleLoadDataExLog (hashcat_ctx, &device_param->cuda_module, binary);
+          const int rc_cuModuleLoadDataEx = hc_cuModuleLoadDataExLog (hashcat_ctx, &device_param->cuda_module, binary, device_param->kernel_threads_max);
 
           if (rc_cuModuleLoadDataEx == -1) return -1;
 
@@ -7853,7 +7864,7 @@ int backend_session_begin (hashcat_ctx_t *hashcat_ctx)
 
         if (device_param->is_cuda == true)
         {
-          const int rc_cuModuleLoadDataEx = hc_cuModuleLoadDataExLog (hashcat_ctx, &device_param->cuda_module, kernel_sources[0]);
+          const int rc_cuModuleLoadDataEx = hc_cuModuleLoadDataExLog (hashcat_ctx, &device_param->cuda_module, kernel_sources[0], device_param->kernel_threads_max);
 
           if (rc_cuModuleLoadDataEx == -1) return -1;
         }
@@ -8021,7 +8032,7 @@ int backend_session_begin (hashcat_ctx_t *hashcat_ctx)
 
             // tbd: check for some useful options
 
-            const int rc_cuModuleLoadDataEx = hc_cuModuleLoadDataExLog (hashcat_ctx, &device_param->cuda_module_mp, binary);
+            const int rc_cuModuleLoadDataEx = hc_cuModuleLoadDataExLog (hashcat_ctx, &device_param->cuda_module_mp, binary, 0);
 
             if (rc_cuModuleLoadDataEx == -1) return -1;
 
@@ -8105,7 +8116,7 @@ int backend_session_begin (hashcat_ctx_t *hashcat_ctx)
 
           if (device_param->is_cuda == true)
           {
-            const int rc_cuModuleLoadDataEx = hc_cuModuleLoadDataExLog (hashcat_ctx, &device_param->cuda_module_mp, kernel_sources[0]);
+            const int rc_cuModuleLoadDataEx = hc_cuModuleLoadDataExLog (hashcat_ctx, &device_param->cuda_module_mp, kernel_sources[0], 0);
 
             if (rc_cuModuleLoadDataEx == -1) return -1;
           }
@@ -8276,7 +8287,7 @@ int backend_session_begin (hashcat_ctx_t *hashcat_ctx)
 
             // tbd: check for some useful options
 
-            const int rc_cuModuleLoadDataEx = hc_cuModuleLoadDataExLog (hashcat_ctx, &device_param->cuda_module_amp, binary);
+            const int rc_cuModuleLoadDataEx = hc_cuModuleLoadDataExLog (hashcat_ctx, &device_param->cuda_module_amp, binary, 0);
 
             if (rc_cuModuleLoadDataEx == -1) return -1;
 
@@ -8360,7 +8371,7 @@ int backend_session_begin (hashcat_ctx_t *hashcat_ctx)
 
           if (device_param->is_cuda == true)
           {
-            const int rc_cuModuleLoadDataEx = hc_cuModuleLoadDataExLog (hashcat_ctx, &device_param->cuda_module_amp, kernel_sources[0]);
+            const int rc_cuModuleLoadDataEx = hc_cuModuleLoadDataExLog (hashcat_ctx, &device_param->cuda_module_amp, kernel_sources[0], 0);
 
             if (rc_cuModuleLoadDataEx == -1) return -1;
           }