A developer shares their first experience setting up ROCm on AMD's Strix Halo system, successfully configuring 128GB of shared memory between CPU and GPU. The setup process involved updating to Ubuntu 24.04 LTS, installing the latest BIOS, and adjusting memory allocation settings through both BIOS and GRUB configurations.

The key challenge was configuring memory sharing, requiring careful adjustment of reserved video memory and GTT settings. The developer discovered that while the GPU can access both reserved and GTT memory, simultaneous utilization can be less efficient due to fragmentation. Some legacy software might see only 512MB of GPU memory, though this wasn't encountered during testing.

After resolving the memory configuration, the developer successfully ran PyTorch with UV package manager and deployed Llama.cpp with the Qwen3.6 model. The setup included custom configurations for both PyTorch dependencies and Opencode integration. Despite some initial rough edges, the system proved capable of handling large context windows and complex AI workloads, demonstrating ROCm's viability for machine learning development on AMD hardware.