Can my NAS run this AI?

Why NPU TOPS don’t matter here

NAS makers now print “AI TOPS” on the box. For a chatbot, that number is close to irrelevant. Ollama, the usual way to run models at home, lists four ways to accelerate a model: NVIDIA CUDA, AMD ROCm, Apple Metal and Vulkan. It does not list a single NPU. The NPU sits idle while you chat.

The reason is how text generation works. To produce each word, the processor reads every active weight of the model out of RAM once. An 8B model at Q4 is 4.9 GB, so 10 words a second means reading about 50 GB a second. Arithmetic is cheap; moving the weights is the bottleneck. That makes memory bandwidth the speed limit and RAM capacity the size limit.

The clearest proof is in one product line. The Minisforum N5 Pro and N5 MAX both carry a 50 TOPS NPU. The N5 Pro reads memory over a 128-bit DDR5 bus at up to 89.6 GB/s; the N5 MAX over a 256-bit LPDDR5x bus at 256 GB/s. Same NPU, and the MAX is roughly three times faster at chat. NPUs do earn their keep elsewhere: photo search in some NAS apps and camera detection in Frigate can use them.

How the calculator works

1. Bandwidth = channels × bus width (bits) ÷ 8 × memory speed (MT/s) ÷ 1000, in GB/s. One DDR5-4800 stick: 1 × 64 ÷ 8 × 4800 ÷ 1000 = 38.4 GB/s. Two: 76.8 GB/s.

2. Memory needed = the model’s actual Ollama download size + the conversation (KV) cache + a runtime allowance of a flat 0.5 GB. The cache is 2 × layers × KV heads × head size × 2 bytes per token of context, from each model’s published configuration, at Ollama’s default f16. Gemma 3 and gpt-oss use sliding-window attention on most layers, which keeps their cache small.

3. Will it fit? Usable RAM = installed RAM − 2 GB for the NAS operating system − RAM used by other apps (4 GB for Plex and Immich, 8 GB for a busy homelab; Immich alone recommends 8 GB of system RAM). Under 15% headroom is “tight”.

4. Speed = bandwidth × efficiency ÷ bytes read per token (the active weights plus the cache). On AMD Radeon and Strix Halo machines efficiency is 0.8, set from published measurements. A mixture-of-experts model only reads its active experts, times a factor of 0.72. For Intel chips with no published measurement we show a range from a CPU-only run (0.15 of bandwidth) up to 0.8. For the Intel N100 and the Rockchip RK3588 we use published measurements on those chips directly, and those disagree, so the range is wide and marked low confidence.

5. Speed bands: under 2 tokens/sec, too slow for chat but fine for overnight batch jobs; 2 to 5, usable but you will wait; 5 to 15, a comfortable reading speed; over 15, faster than you can read.

How close are the estimates?

Every published measurement we hold is re-run through the engine on each build, and the build fails if an estimate drifts more than 15% from the measurement. Two honest caveats. The 0.8 efficiency and the 0.72 mixture-of-experts factor were set from the Ryzen AI Max+ 395 rows, so matching those rows proves nothing on its own; the independent checks are the HX 370 rows and the gpt-oss-20b rows. And the mixture-of-experts factor rests on a single measurement (Qwen3 30B-A3B). A second model, gpt-oss-20b, lands within 9% on two machines but 22% low against one Vulkan run on the 395, so mixture-of-experts results show a wider ±40% range.

Frequently asked questions