Add RDNA4 GPU support (RX 9070 XT, RX 9070, RX 9060 XT, RX 9000 XT) #1033

Target release

Unsure

Optimization domain

Both

Problem / use case

RDNA4 GPUs (Radeon RX 9000 series) are not currently recognized. These consumer GPUs use WMMA matrix instructions with native FP8 support and 16GB GDDR6 at 640 GB/s, requiring Windows ROCm 7.3+ support via hipConfig --show-device autodetection.

Proposed solution

Add RDNA4 GPU types (rx9070xt, rx9070, rx9060xt, rx9000xt) to gpu_types.py
Add RDNA4 HW specs to roofline_ceiling.py (16GB GDDR6, 640 GB/s BW, 389 TFLOPS FP8 peak)
Add Windows-compatible fcntl handling and os.getuid guards
Add POSIX-only test guards for cross-platform compatibility

Expected impact

Enables Hyperloom to run on consumer RDNA4 GPUs for model inference optimization on Windows 11.

Alternatives considered

No response

Additional context

No response