Comprehensive investigation of RX 6800 (gfx1030) stability and performance: Linux KFD crashes vs. Windows ROCm limits

### Custom Node Testing

- [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Your question

Hello,

I have conducted a detailed 4-day technical investigation regarding local image generation using an AMD Radeon RX 6800 (gfx1030 / RDNA2) on both Linux and Windows.

I wanted to share my findings with the repository maintainers and the community, as it sheds light on driver stability and the fundamental limits of high-resolution generation on this specific GPU architecture.

### Summary of findings:
1. **Linux (Ubuntu, amdgpu/ROCm):** Experienced repeated `[gfxhub] Page fault observed @ 0x0000000000000000` crashes causing GPU ring timeouts. The investigation tracked this down to a KFD SVM/HMM subsystem bug (`svm_range_restore_work` hogging the CPU), which is an upstream driver issue, not an application-level bug. The only workaround was using `--cpu-vae`.
2. **Windows 11 (Native ROCm nightly / TheRock):** The setup is highly stable up to ~1M pixels (1024x1024). However, at 2048x2048 it either hard-crashes the process or experiences a 58x slowdown. This is caused by the mandatory `MATH SDP` attention backend forced for RDNA1/2 cards, which scales quadratically $O(n^2)$ relative to the token count.

### Full detailed report
I have documented the complete timeline, coredump analysis, hypotheses tested, and performance metrics in the following GitHub Gist:
https://gist.github.com/AMD-AI-Enthusiast/e695189b5360c0616d8e1a4176665585

Hopefully, this data helps other users trying to optimize their RDNA2 setups or assists developers in understanding memory constraints with legacy AMD cards. Thank you for maintaining this project!

### Logs

```powershell

```

### Other

_No response_

Start by reading the linked GitHub Gist and comparing its RX 6800 findings for Linux KFD crashes, Windows ROCm limits, and the --cpu-vae workaround. The issue names no repository file, test, or requested code change, so reproduction and a definition of done would need to be established with maintainers.

Written by the indexing model from the issue text.