Recently, I ran Qwen3.6-35B-A3B on 2x RX9060 16GB vram each + 32GBs RAM with rocm. I felt that ROCM had an "I bought the whole piano I'm gonna use the whole piano" vibe; it pushes the GPUs to its limit as if there were no tomorrow.

The average power consumption is 140W in both GPUs, with temperatures rising to +20°C during extended runs. Ultimately, I didn't notice a significant performance difference compared to Vulkan, which averages 90W and doesn't put much strain on the GPUs.

Meanwhile, I'll be using Vulkan, but checking different configurations and parameters weekly with ROCM.

Vulkan is working perfectly with this model. Using harness engineering, I can implement complex features in real-world projects.