PSA: BF16 is faster than int8-convrot on DGX Spark
retained text
I had been using ZIT bf16 since before the convrot update to Comfy. Today I finally decided to see today how much faster convrot would make it, I was honestly expecting it to be 40-50% faster. **ZIT Test** 1. Open Comfy's built-in "Z-Image-Turbo Int8: Text to image" template 2. Choose a model 3. Run the workflow once to warm up the model cache 4. Change the first character of the prompt to break text prompt cache 5. Run and measure Result: - z_image_turbo_int8_convrot.safetensors = 6.3GB = 6.60s - z_image_turbo_bf16.safetensors = 13GB = 7.63s - z_image_turbo_nvfp4.safetensors = 4.6GB = 5.37s I repeated the test 3 times because I thought I must be messing something up. Nope, identical results. So it's just 1s difference between int8 and bf16? That can't be right. I figured ZIT is just too fast for int8's benefits to show. That I would need a bigger model to really see the difference. How about the biggest, slowest and bestest of them all, then? **H3 test** - Open Comfy's built-in H3 r2v template, delete the turbo LoRA node (didn't have it), delete the 2nd image ref node, set the remaining reference to the built-in example.png 768x768 mspaint blonde girl - Leave everything else at default, ie 0.4MP and 5.0s - Warm up the cache, change 1st letter of prompt, only measure the 2nd gen, repeated twice to believe my eyes Result: - minimax_h3_ref2va_pruned_int8_convrot.safetensors = 21GB = **286.66s for 5s video** - minimax_h3_ref2va_pruned_bf16.safetensors = 40.2GB = **272.22s for 5s video** That's right. bf16 is actually faster. What's going on? What are we doing here? My launch settings, for completion's sake. I'm too lazy to repeat the test without ck attention, I don't even know if it has any effect, I'm just a casual. `--use-ck-attention --disable-mmap --cache-classic`.
document fetched 2026-09-30 00:35 · canonical https://reddit.com/r/StableDiffusion/comments/1wszbe0/psa_bf16_is_faster_than_int8convrot_on_dgx_spark