Can I Run OpenAI gpt-oss 120B on AMD Strix Halo 128GB? Unified Memory Benchmark
Model weights, initial KV-cache, and CUDA context buffers.
Available safety margin for OS compositing and context expansion.
Throughput determined by memory bus bandwidth.
Hardware Physics & Bottleneck Analysis
Holding 116.8 billion parameters requires an enormous unified memory space. While traditional desktops split memory across two PCIe slots, AMD Strix Halo places 128GB of coherent LPDDR5X-8533 memory directly on the APU substrate (256 GB/s). ROCm and Vulkan can allocate up to 96GB to the integrated GPU, leaving 32GB for the operating system and IDE.
On Linux, configure the ROCm GTT memory ceiling via kernel boot parameter `amdgpu.gttsize=98304` to grant llama.cpp access to the full 96GB pool, and run `llama-server -fa on -ctk q8_0` for maximum prompt-processing throughput.
Buy Hardware on Amazon
Direct Prime-badged hardware accessories and GPUs with dynamic regional routing.
Frequently Asked Questions
Why does Dual RTX 3090 48GB fail on gpt-oss 120B?
Dual RTX 3090 provides 48GB VRAM total. At Q4_K_M, gpt-oss 120B requires 71.17 GB—exceeding 48GB by 23.17 GB. Offloading those 23 GB to DDR5 system RAM over PCIe cuts speed from 35 tok/s down to 1.8 tok/s.
Does Strix Halo 128GB get noisy when running 120B models?
No. The total system TDP is only 85W–120W, operating whisper-quiet in compact mini-PC chassis compared to 750W dual-GPU desktop rigs.
Calculate Exact Context Window Headroom
Explore live token throughput and VRAM allocation sliders in the interactive matchmaker.
Launch Live Matchmaker Calculator →