Run private chatbots, coding assistants and image models at home. For LLMs, memory size and bandwidth matter far more than the NPU's TOPS figure — from 7B models on a Radeon 780M to 70B+ on a 128 GB Strix Halo.
A 4-bit quantised model needs roughly 0.6 GB per billion parameters plus context: ~5 GB for 8B, ~10 GB for 14B, ~20 GB for 32B and ~42 GB for 70B. Buy the RAM for the largest model you want to run.
Token generation is bandwidth-bound. Dual-channel DDR5 (~90 GB/s) is fine for 7B–14B models; Strix Halo's 256-bit bus (~256 GB/s) is what makes 30B–70B models usable.
Ollama, llama.cpp and LM Studio run on the iGPU via ROCm or Vulkan. NPUs (XDNA 2, Intel NPU) are useful for smaller vendor-supported models and Windows AI features, but rarely for general LLM inference today.
AMD iGPUs often need an HSA_OVERRIDE_GFX_VERSION tweak for ROCm or a Vulkan backend; Intel uses OpenVINO or IPEX-LLM. Check your toolchain supports the exact GPU.
Radeon 780M with 32–64 GB DDR5 runs quantised 7B–14B models at conversational speeds — see our Ollama tutorial.
| 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 |
Native 1080p AAA gaming, external eGPU docking, Ollama local 7B-13B LLM inference.
Radeon 890M and up to 128 GB of SODIMM RAM fit larger models; OCuLink adds a discrete-GPU path.
| 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 |
Local LLM inference, 1080p–1440p gaming with upscaling, software builds and creative workloads in a compact desktop.
Arc B390 with XMX and a 50 TOPS NPU — strong on OpenVINO-optimised models with very low idle power.
| Graphics (iGPU) | Intel Arc B390 (12 Xe3-cores) |
| Memory Architecture | LPDDR5X-9600 soldered (X-series) or DDR5-7200 SODIMM |
| Storage Exp. | 2x M.2 PCIe 4.0 / 5.0 (chassis dependent) |
| Idle Power | 5W – 10W |
| Est. Barebone Cost | $850 - $1,600 |
Quick Sync media servers and editing, efficient always-on workstations, 1080p gaming on the iGPU, on-device AI with OpenVINO.
Up to 128 GB unified memory on a 256-bit bus — the only mini PC that holds 70B-class models entirely in GPU-addressable RAM.
| Graphics (iGPU) | Radeon 8060S (40 CUs, up to 2.9 GHz) |
| Memory Architecture | LPDDR5X-8000 256-bit, soldered (64GB or 128GB) |
| Storage Exp. | 2x M.2 PCIe 4.0 x4 |
| Idle Power | 10W – 18W |
| Est. Barebone Cost | $1,500 - $2,800 |
Local LLM inference on 30B–120B models, ROCm/AI development, 1440p gaming without a discrete GPU, heavy multi-core builds.
| Memory | 32 GB minimum; 64 GB for 32B models; 128 GB for 70B+ |
| Graphics | Radeon 780M / 890M / 8060S or Arc B390 |
| Bandwidth | Dual-channel DDR5-5600+ or LPDDR5X; 256-bit for large models |
| Storage | 1 TB+ NVMe — model files are 5–80 GB each |
| Expansion | OCuLink if you may add a discrete GPU with dedicated VRAM |
| Graphics | AMD Radeon 780M (12 CUs, RDNA 3, up to 2.7 GHz) |
| Memory | 2x DDR5-5600 SODIMM (Up to 128GB) |
| Storage | 2x M.2 2280 PCIe 4.0 (Up to 16TB) |
| Idle Power | 13W – 15W |
| Price | $409.99 barebone · $919.99 (32GB/1TB) |
Ryzen 7 8845HS / 7840HS Beast
| Graphics | AMD Radeon 890M (16 CUs, RDNA 3.5) |
| Memory | 2x DDR5-5600 SODIMM (Up to 128GB) |
| Storage | 3x M.2 2280 PCIe 4.0 (Up to 12TB) |
| Idle Power | 5.5W – 9W |
| Price | $719 barebone · €1,399 (32GB/1TB) |
Gorgon Point Ryzen AI 9 HX 470
| Graphics | AMD Radeon 890M (16 CUs, RDNA 3.5, 3.1 GHz) |
| Memory | 2x DDR5-5600 SODIMM (32GB or 64GB) |
| Storage | 2x M.2 2280 PCIe 4.0 x4 (Up to 8TB) |
| Idle Power | ~8W idle (Robtech, via Starry Hope) |
| Price | $1,399 (32GB/1TB) · $1,699 (64GB/1TB) |
Gorgon Point Ryzen AI 9 HX 470
| Graphics | Intel Arc B390 (12 Xe cores, 2.5 GHz, 122 TOPS INT8) |
| Memory | LPDDR5X-8533 onboard (64GB) |
| Storage | 2x M.2 2280 (PCIe 5.0 x4 + PCIe 4.0 x1, up to 8TB each) |
| Idle Power | Not published |
| Price | $2,399.99 (64GB/1TB) |
Intel Panther Lake Core Ultra X9 388H / X7 358H