GEEKOM IT15 (Ultra 9 285H)

The GEEKOM IT15 (Ultra 9 285H) is a NUC-class mini PC that packs Intel’s Arrow Lake-H flagship mobile silicon into a 4.4” x 4.6” chassis. At $1,399 (MSRP, though street pricing often lands ~$1,200), it targets the prosumer edge of the AI inference market—a bridge between consumer-grade laptops and workstation-class machines. GEEKOM has been a reliable builder of compact systems, and with this model they’ve pushed the combined platform AI performance to 99 TOPS, leveraging the CPU, Intel Arc 140T iGPU, and NPU simultaneously.

For practitioners running 8B-class models locally, this hardware hits a practical sweet spot. It’s not a data-center GPU, but it doesn’t need to be; the shared 32GB DDR5 system memory (90 GB/s bandwidth) acts as unified VRAM for the iGPU, enabling 8B parameter models at Q4–Q5 quantizations with workable token rates. The 65W sustained TDP (peaking around 100W briefly) means it’s viable for always-on inference servers, field deployments, or desktop AI workstations that don’t demand a rack.

What sets the IT15 apart is its efficiency per watt. You can run local LLMs, multimodal models, and agent pipelines without a dedicated GPU, at a power draw that won’t spike your utility bill. It competes directly with high-end mini PCs from Minisforum and ASUS (e.g., NUC 14 Pro+), and with lower-TDP laptops that don’t have the same memory bandwidth.

• 32GB DDR5-5600 dual-channel, shared with the Intel Arc 140T iGPU.
• Effective VRAM: ~8 GB available to the GPU after OS and background processes (Windows 11 Pro). This is the hard ceiling for model weight + context + KV cache.
• Memory bandwidth: 90 GB/s. For comparison, a desktop RTX 4060 has ~272 GB/s; the IT15’s bandwidth is the primary bottleneck for token generation speed.

• Combined platform AI: 99 TOPS (INT8) across CPU + Arc iGPU + NPU. In practice, most inference runs on the iGPU (~70 TOPS) with CPU fallback for memory-bound operations.
• FP16 performance on the iGPU is roughly 2.7 TFLOPS (estimated via INT8/FP16 ratio), adequate for simple vision tasks but not for training.
• NPU: Intel AI Boost handles lightweight, always-on audio and vision preprocessing (e.g., wake-word detection, background blur). Not used for LLM inference.

• Sustained TDP: ~50-55W after boost expires (≈30s burst up to 90-100W). Real-world power draw under LLM inference load is ~45-55W, making it energy-efficient for 24/7 operation.
• Thermals: Fan noise is audible under sustained load (reviewers note it as “loud”). The chassis stays cool to the touch, but the system throttles to maintain safe temps. For AI workloads, you’ll want good airflow or a quiet location.

• 2x USB4 (40 Gbps) can be used for external GPUs (eGPU), though OCuLink is absent—a notable gap for those wanting to add a dedicated card later.
• 2.5GbE LAN, WiFi 7, BT 5.4, dual HDMI 2.1, SD card slot. Supports 8K quad display output.
• Storage: 2TB NVMe PCIe Gen4 (tested sequential writes ~1.2 GB/s when almost full—adequate for model storage and swap).

The shared memory architecture means model fitting depends on both weight size and context length. Here’s what works:

• Llama 3.1 8B (Q4_K_M): ~5.5 GB weights + 2 GB buffer for 4K context. Token generation: ~10–15 tok/s on iGPU with llama.cpp (CPU offload layer mixing). Acceptable for chat, summarization, code generation.
• Mistral 7B (Q5_K_M): ~4.8 GB, runs at 15–18 tok/s. Excellent for real-time conversation.
• Qwen 2.5 7B (Q4): Similar performance. Multilingual tasks are smooth.
• DeepSeek-R1 7B (Q4): Runs well; CoT reasoning may stress context window—keep under 8K tokens for best speed.

• Llama 3.1 13B (Q3_K_M): ~7 GB weights. Leaves minimal headroom for context. Expect 6–9 tok/s, with frequent KV cache evictions. Fine for batch inference or async tasks.
• Mixtral 8x7B (Q2) : Too large (~10 GB) – will exceed VRAM and swap to system RAM, dropping to <2 tok/s. Not recommended.

• LLaVA 1.5 7B (Q4): Runs at ~10 tok/s for text generation; image encoding uses iGPU compute. Works for simple image captioning.
• Qwen-VL 7B: Similar behavior. Processing large images (4K) may cause 1–2s latency per image.

• With 8B Q4 models, you can push to 8K–16K context before hitting the wall. Beyond that, use sliding window or Flash Attention. The 90 GB/s bandwidth limits speed—expect 40–50% slower than a desktop RTX 4070.

8B at Q4_K_S with 4K context. This yields ~13–15 tok/s, balancing quality and speed. For coding or structured tasks, Q8_0 on a 3B model (e.g., Phi-3-mini) runs at 25+ tok/s—great for autocomplete and agent scripts.

If you want a dedicated always-on machine for ChatGPT-style queries, local RAG, or proof-of-concept agents, the IT15 eliminates GPU complexity. Plug in, install LM Studio or Ollama, and start inferring. The 2TB SSD holds dozens of quantized models.

The 65W sustained power, compact size (less than 2” tall), and passive capability under light load make it viable for remote edge nodes—digital signage with AI, on-device transcription, or low-volume inference at a retail location. Pair with a battery UPS for mobile use.

Test agents, chaining calls to local models, or run a local inference server for a small team (1–3 concurrent users). The USB4 port supports an external GPU if you later need more VRAM. The IT15 is a flexible dev sandbox.

The shared memory and iGPU compute (2.7 TFLOPS FP16) are insufficient for any meaningful training or fine-tuning. Stick to inference.

• Price: MS-01 with RTX 4060 ~$1,500.
• VRAM: 8GB dedicated GDDR6 vs. shared 8GB DDR5. The RTX 4060 gives ~220 GB/s bandwidth, enabling 8B Q4 models at 25–30 tok/s—roughly 2x faster than the IT15.
• Form factor: MS-01 is slightly larger but includes OCuLink for future eGPU.
• When to pick IT15: Lower power draw (65W vs 120W+), smaller footprint, quieter under light load, and similar VRAM ceiling. Choose the IT15 for energy-sensitive edge or desktop where speed isn’t critical.

• CPU: The 285H is Arrow Lake (15th Gen), slightly faster single-core and better NPU. The 185H is Meteor Lake (14th Gen)
• Memory bandwidth: Both 90 GB/s. The NUC 14 Pro+ uses LPDDR5x, achieving similar numbers.
• Price: NUC 14 Pro+ ~$1,100, cheaper but with 1TB SSD and sometimes 16GB RAM.
• When to pick IT15: Comes fully loaded (32GB + 2TB) out of box. The 285H’s NPU improvement matters if you use Windows Studio Effects or other ONNX runtime tasks. If budget is tighter, the NUC 14 Pro+ is a solid alternative with comparable inference performance.