Unified-Memory AI Workstation

Strix Halo Ryzen AI Max+ 395

AMD's big APU: 16 full Zen 5 cores, a 40-CU Radeon 8060S and a 256-bit LPDDR5X bus with up to 128 GB of unified memory — the only mini PC chip that can hold 70B-class LLMs entirely in GPU-addressable RAM.

Recommended for Gaming Local AI & LLMs NAS & Storage

ProcessorAMD Zen 5 (16C/32T, 50 TOPS NPU)

GraphicsRadeon 8060S (40 CUs, up to 2.9 GHz)

Idle Power10W – 18W

Est. Barebone$1,500 - $2,800

Mini PCs with this chip

No mini PCs listed in this category yet.

Full Specifications

Processor

Model	AMD Ryzen AI Max+ 395 (Strix Halo)
Cores / Threads	16 / 32 (all full Zen 5, two CCDs)
Boost	Up to 5.1 GHz
L3 Cache	64 MB (plus 32 MB Infinity Cache for the iGPU)
Default TDP	55 W (configurable 45–120 W; mini PCs typically run 120–140 W)
NPU	AMD XDNA 2, up to 50 TOPS (126 TOPS platform total)
Process	TSMC 4 nm (CCDs) + 4 nm I/O-GPU die

Graphics & Media

iGPU	Radeon 8060S, 40 CUs RDNA 3.5, up to 2.9 GHz
Media Engine	H.264 / HEVC / AV1 encode and decode
Display Output	Up to 4 displays, single display up to 8K
GPU Memory	Up to 96 GB dedicated in Windows; ~110 GB+ via GTT on Linux

Memory & Storage

Memory Channels	Quad-channel 256-bit (~256 GB/s theoretical bandwidth)
Memory Type	LPDDR5X-8000, soldered — not upgradeable
Max Memory	128 GB
Storage	PCIe 4.0 x4 NVMe — usually 2x M.2 2280

Connectivity & Power

USB4	Up to 2x 40 Gbps
PCIe	16 lanes PCIe 4.0 total (no OCuLink on most designs)
Ethernet	2.5GbE to 10GbE depending on model
Typical Idle	10 – 18 W at the wall; 150–200 W under full CPU + GPU load

Strengths

Up to 128 GB unified memory — runs models no 24 GB consumer GPU can fit
Radeon 8060S performs close to an RTX 4060 laptop GPU
Sixteen full Zen 5 cores: desktop-class multi-core in a small box
Quad-channel memory bandwidth roughly double any other mini PC chip

Trade-offs

Memory is soldered — buy the capacity you will need on day one
Expensive, and 128 GB configurations carry a large premium
~256 GB/s still limits token speed on large dense models versus discrete GPUs
High sustained power means larger chassis and audible fans

Best Suited For

Local LLM inference on 30B–120B models, ROCm/AI development, 1440p gaming without a discrete GPU, heavy multi-core builds.

Check before you buy

Choose 128 GB if local LLMs are the goal — 64 GB rules out 70B models at useful quantisation.
On Linux, raise the GTT/TTM limit (or set UMA in BIOS) so the GPU can address most of the RAM.
Check the sustained power mode: 120 W+ chassis are markedly faster than 55–85 W designs.
Confirm ROCm support for your toolchain (gfx1151) or plan on Vulkan backends like llama.cpp.

Compare with

High-Performance & eGPU

Ryzen 7 8845HS / 7840HS Beast

AMD Zen 4 Hawk Point (8C/16T, 5.1GHz)

Graphics (iGPU)	Radeon 780M (12 CUs RDNA 3)
Memory Architecture	Dual DDR5 5600MHz SODIMM
Storage Exp.	2x PCIe 4.0 NVMe (Raid 0/1 support)
Idle Power	11W – 16W
Est. Barebone Cost	$490 - $680

Best Suited For

Native 1080p AAA gaming, external eGPU docking, Ollama local 7B-13B LLM inference.

Flagship AI & Workstation

Strix Point Ryzen AI 9 HX 370

AMD Zen 5 (12C/24T, 50 TOPS NPU)

Graphics (iGPU)	Radeon 890M (16 CUs)
Memory Architecture	LPDDR5X 7500MT/s (Soldered 32/64GB)
Storage Exp.	Dual M.2 PCIe 4.0 x4
Idle Power	12W – 19W
Est. Barebone Cost	$790 - $1,050

Best Suited For

Ultra-fast local model fine-tuning, 1440p gaming, intensive creative studio workflow.

Flagship AI & Workstation

Gorgon Point Ryzen AI 9 HX 470

AMD Zen 5 (12C/24T, 55 TOPS NPU)

Graphics (iGPU)	Radeon 890M (16 CUs, up to 3.1 GHz)
Memory Architecture	DDR5-5600 SODIMM or LPDDR5X (Up to 128GB)
Storage Exp.	2–3x M.2 PCIe 4.0 x4 (chassis dependent)
Idle Power	6W – 10W
Est. Barebone Cost	$720 - $1,400

Best Suited For

Local LLM inference, 1080p–1440p gaming with upscaling, software builds and creative workloads in a compact desktop.