fp16 is actually faster than fp32 on a GTX 1080.

comfy/model_management.py

@@ -987,16 +987,13 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
     if props.major < 6:
         return False
 
-    fp16_works = False
-    #FP16 is confirmed working on a 1080 (GP104) but it's a bit slower than FP32 so it should only be enabled
-    #when the model doesn't actually fit on the card
-    #TODO: actually test if GP106 and others have the same type of behavior
+    #FP16 is confirmed working on a 1080 (GP104) and on latest pytorch actually seems faster than fp32
     nvidia_10_series = ["1080", "1070", "titan x", "p3000", "p3200", "p4000", "p4200", "p5000", "p5200", "p6000", "1060", "1050", "p40", "p100", "p6", "p4"]
     for x in nvidia_10_series:
         if x in props.name.lower():
-            fp16_works = True
+            return True
 
-    if fp16_works or manual_cast:
+    if manual_cast:
         free_model_memory = maximum_vram_for_weights(device)
         if (not prioritize_performance) or model_params * 4 > free_model_memory:
             return True