add the host-only row: 646.1 s against 60.2 on the card, same machine and seed

README.md

@@ -243,8 +243,10 @@ every run:
 | 512×288 | 2.5 s | 34.8 s | 16.6 s | 4.3 s | **60.2 s** |
 | 512×256 | 2.3 s | 31.4 s | 8.4 s | 4.3 s | **48.5 s** |
 | 256×160, 22 frames | | | | | **15.9 s** |
+| 512×288, host only (`CMF_MMH3_GPU=0`) | 2.6 s | 363.5 s | 271.5 s | 7.1 s | **646.1 s** |
 
-At 8 steps a 512×288 clip is **91.6 s**; at 2 it is 42.8. The same 4-step
+The card is **10.7× the host path** on the same machine, same seed. At 8
+steps a 512×288 clip is **91.6 s**; at 2 it is 42.8. The same 4-step
 render took 172 s when this card was written — the pipeline has since moved
 onto the card end to end, both VAE decoders with it.