AMD Strix Halo (Ryzen AI Max+ 395)

AMD Strix Halo (retail name Ryzen AI Max+ 395) launched in 2025 as AMD’s answer to Apple’s unified-memory inference story. A Zen 5 CPU combined with a 40-CU RDNA 3.5 iGPU and an XDNA 2 NPU, all sharing up to 128 GB of soldered LPDDR5X-8000 memory at 256 GB/s. AMD’s path here stands on decades of x86 work plus the Radeon RDNA lineage (RDNA 1 on the 5700 XT in 2019) and the Xilinx-acquired NPU lineage from the XDNA team.

Who it’s for: plebs who want one box that does development, inference, and general computing without the fan noise of a GPU rig. Mini-PC form factor (Framework Desktop, HP ZBook variants) and high-end laptops (ASUS ROG Flow, Razer Blade).

Models it runs comfortably: with 128 GB unified, Llama 3 70B at Q8, Mixtral 8x22B at Q4, Qwen 2.5 72B at Q5_K_M. ROCm and llama.cpp Vulkan backend are the practical runners on Linux; DirectML and ONNX Runtime on Windows. Expect 5–15 tok/s on 70B-class models — slower than a 4090 but the VRAM ceiling is ~5× higher.

Hashcenter notes: fits in sub-200W laptop/mini envelope (configurable TDP 45–120 W). Completely silent or near-silent in most chassis. Runs on USB-C PD or standard barrel adapters. Credit to AMD for bringing unified-memory inference to the x86/Linux ecosystem, and to the ROCm and llama.cpp communities for making the software stack usable.

Further reading: This card is a core component of a pleb-grade AI Hashcenter. Pair it with the sovereignty argument in the Sovereign AI for Bitcoiners Manifesto, or look at how the same 120V envelope powers a Bitcoin space heater in our mining catalog. Running both workloads on one rig? See Heating Your Home With Inference.