Add bfloat16

src/utils.py

@@ -223,7 +223,7 @@ def get_peak_flops(gpu_name, precision):
 def transfer_precision2bytes(precision):
     if precision == "float32":
         return 4
-    elif precision == "float16":
+    elif precision in ["float16", "bfloat16"]:
         return 2
     elif precision == "8bit":
         return 1