What you will be able to do
- Compare ready-made devices by the two numbers from lesson 9.2.
- Decide when buying a device beats building a PC.
- Recognize the trade-offs of each option without reading marketing material.
Why devices are different
A graphics card puts a small amount of very fast memory next to its processor. A unified-memory machine puts a large amount of moderately fast memory where both the processor and the graphics can use it. The first wins on speed for models that fit; the second wins on size. Apple Silicon Macs pioneered this design for personal machines; AMD's Ryzen AI Max+ chips and NVIDIA's DGX Spark follow it.
The comparison
| Device | Memory | Bandwidth | Price (CAD) | Possible use to test | Trade-off | Source |
|---|---|---|---|---|---|---|
| Mac mini (M6) | 16 GB unified (24 or 32 GB optional) | 153 GB/s (170 GB/s with 24 or 32 GB) | from CA$1,249 | Trying local AI quietly on a desk; small models. | Memory cannot be added later. Small models only at 16 GB. | [1] [2] |
| Mac mini (M5 Pro) | 24 GB unified (up to 64 GB) | 307 GB/s | from CA$2,399 | Mid-size models in a small, silent box. | Bandwidth is a third of a fast graphics card; 64 GB ceiling. | [1] [2] |
| Mac Studio (M5 Max) | 36 GB unified (48, 64 or 128 GB optional) | 460 GB/s (614 GB/s with the 40-core GPU) | from CA$3,499 | Large models and long documents on one quiet machine. | Price climbs steeply with memory; no upgrades after purchase. | [1] [2] |
| Mac Studio (M5 Ultra) | 96 GB unified (256 GB optional; 512 GB announced for late October2026) | 1.2 TB/s | from CA$7,799 | A large unified-memory option; test model fit and speed. | Very expensive at the top configurations. | [1] [2] |
| Mini PC or workstation with AMD Ryzen AI Max+ 395 | up to 128 GB LPDDR5X, shared with the built-in Radeon 8060S graphics | about 256 GB/s (derived: 8,000 MT/s × 256-bit ÷ 8) | varies by maker (not verified) | A large memory pool for MoE models on Windows or Linux. | Bandwidth is modest, so dense models run slowly; software support is younger than NVIDIA’s. | [1] |
| NVIDIA DGX Spark | 128 GB LPDDR5X unified; 64 GB is OEM-specific | 273 GB/s | not verified | NVIDIA’s software stack in a desktop box with a large memory pool. | Bandwidth similar to the AMD option; runs Linux; priced as a developer workstation. | [1] |
| Desktop PC with a 32 GB graphics card | 32 GB on the card plus system RAM | 640 to 1,792 GB/s on the card | see the three builds in lesson 9.4 | A discrete-GPU option; compare measured results and current costs. | Models larger than the card spill into slow system RAM; louder and more power-hungry. | [1] [2] |
| Used workstation or server with several GPUs | sum of the cards (e.g. 2 × 24 GB) | per card; layers split across cards | varies widely (not verified) | Experienced builders who want large VRAM pools cheaply. | Power, noise, heat, and no warranty. Splitting a model across cards adds setup work. |
As of 16 October 2026. Specifications and Apple prices from the manufacturer pages linked in each row. Apple prices are apple.ca list prices before tax. “Not verified” marks a price not independently established here; it does not imply that no price exists. Configuration and availability must be checked before buying.
Reading the table with the formulas
Take a 63.39 GB model such as gpt-oss-120b. It fits on a Mac Studio with 128 GB, on an AMD Ryzen AI Max+ 395 system with 128 GB, or on a DGX Spark with 128 GB, but not on any single consumer graphics card. [source] [source] [source] Because it is a mixture-of-experts model, only about 2.77 GB is read per token, so even modest bandwidth gives usable speed: 273 GB/s ÷ 2.77 ≈ 99 tokens per second ceiling on the DGX Spark; 614 GB/s ÷ 2.77 ≈ 222 on the faster Mac Studio M5 Max.
Now take a 16.82 GB dense model such as Qwen3.6-27B. All the devices hold it easily, but speed follows bandwidth directly: about 16 tokens per second on a 273 GB/s box, about 36 on a 614 GB/s Mac Studio, and about 107 on a 1,792 GB/s RTX 5090. [source]
The pattern: unified-memory devices are the simplest way to run big MoE models quietly; a desktop graphics card is the fastest way to run dense models that fit in 32 GB.
When to buy rather than build
Buy when you want a quiet, warranty-backed machine with no assembly, when you mainly want big MoE models, or when the machine will sit in an office. Build when you want the fastest answers from mid-size dense models, the option to add a second card later, or the lowest cost per unit of bandwidth. Avoid used multi-GPU servers unless someone on hand has run one before: power, noise, heat, and setup work are real.
Self-check
- What is the main advantage of a unified-memory device, and the main disadvantage?
- Why does gpt-oss-120b run acceptably on a 273 GB/s device while a 27B dense model runs slowly on it?
- Using the table, which device gives the fastest predicted writing speed for a 16.82 GB dense model?
- Name two reasons to buy a device instead of building.
Check your understanding
- Advantage: a large pool so big models fit. Disadvantage: lower bandwidth than a fast graphics card, so dense models write more slowly.
- gpt-oss-120b is mixture-of-experts: only about 2.77 GB is read per token. The dense 27B model reads its full 16.82 GB per token.
- A desktop with an RTX 5090 (1,792 GB/s), at about 107 tokens per second.
- Any two of: quiet, warranty, no assembly, big MoE models, office-friendly.