Skip to content

Native INT8 ConvRot significantly slower than FP8 on A100 (Torch 2.10 + cu130) #14824

Description

@muxuezzz

Custom Node Testing

Your question

Problem

On A100 GPUs, the newly added native INT8 ConvRot models (introduced in v0.27.0) are much slower than FP8 (only 1/3 to 1/2 the speed of FP8), contrary to the claimed 1.5-2x+ speedup on most NVIDIA GPUs.

Tested models:

  • Flux.2-Klein
  • Qwen-Image-Edit (qwenedit)

Environment:

  • GPU: NVIDIA A100 (80GB)
  • PyTorch: 2.10 + cu130
  • ComfyUI: v0.27.0)
  • CUDA: 13.0

Expected Behavior

According to the announcement and community benchmarks on RTX 20/30/40/50 series, INT8 ConvRot should provide:

  • Better or comparable quality to FP8
  • Significantly faster inference (especially on cards with strong INT8 Tensor Core support)

A100 has excellent INT8 Tensor Core performance, so it should benefit as well or even more.

Actual Behavior

  • INT8 ConvRot generation speed is only 33%~50% of FP8.
  • This matches reports from other users that A100 sees little to no speedup (or regression) with the current INT8 implementation.

Additional Context

  • ConvRot is a QuaRot variant using Hadamard rotation + row-wise INT8 to suppress outliers.
  • The custom node No any speed increase BobJohnson24/ComfyUI-INT8-Fast#71 also notes limited/no benefit on A100.
  • The issue is likely due to kernel fusion, dispatch paths, or missing optimizations for Ampere datacenter GPUs (vs consumer RTX cards) in the current native implementation.

Steps to Reproduce

  1. Load an official INT8 ConvRot model (e.g. from Comfy-Org org on HF or modelscope).
  2. Compare with the FP8 version of the same model using the same workflow.
  3. Measure steps/second or total generation time.

Suggested Fix / Help Wanted

  • Improve kernel support / Triton kernels for A100/Ampere in native INT8 path.
  • Better fallback or mixed-precision handling for datacenter GPUs.
  • Profiling data or torch.compile integration improvements for ConvRot operations.

Would be great if the team could add A100-specific benchmarks or optimizations in future releases. Happy to provide more logs, profiler outputs, or test workflows.

Logs

Other

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    User SupportA user needs help with something, probably not a bug.

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions