What you will be able to do
- Choose a graphics card for a budget using the two numbers from lesson 9.2.
- Check a parts list for compatibility before buying.
- Follow the assembly order, first boot, BIOS settings, and driver installation.
Start from the card
Choose the graphics card first, by memory and then bandwidth, and build everything else around it. The table lists current cards from the manufacturers' own specification pages.
| Card | Memory | Bus | Bandwidth | Board power / recommended PSU | Source |
|---|---|---|---|---|---|
| NVIDIA GeForce RTX 5090 | 32 GB GDDR7 | 512-bit | 1,792 GB/s | 575 W / 1000 W | [1] |
| NVIDIA GeForce RTX 5080 | 16 GB GDDR7 | 256-bit | 960 GB/s | 360 W / 850 W | [1] |
| NVIDIA GeForce RTX 5070 Ti | 16 GB GDDR7 | 256-bit | 896 GB/s | 300 W / 750 W | [1] |
| NVIDIA GeForce RTX 5070 | 12 GB GDDR7 | 192-bit | 672 GB/s | 250 W / 650 W | [1] |
| NVIDIA GeForce RTX 5060 Ti 16 GB | 16 GB GDDR7 | 128-bit | 448 GB/s | 180 W / 600 W | [1] |
| AMD Radeon AI PRO R9700 | 32 GB GDDR6 | 256-bit | 640 GB/s | 300 W / 750 W | [1] |
As of 16 October 2026. Displayed rows were checked against the manufacturer specification pages; the two other AMD rows are withheld pending a complete primary check: AMD · Radeon RX 9060 XT specifications, AMD · Radeon RX 9070 XT specifications, AMD · Radeon AI PRO R9700 specifications, NVIDIA · Compare GeForce graphics cards. Prices are not listed by the manufacturers and are deliberately left out.
Three budgets, worked through
Each build below names its card, the rest of the parts by class, what it can run, what the speed formula predicts, and its limits.
Build A: the capable starter
about CA$2,000 to CA$2,800 (estimate)
- Graphics card
- NVIDIA GeForce RTX 5060 Ti 16 GB (448 GB/s)
- Other parts
- Mid-range 6- or 8-core desktop CPU; B-series motherboard with one PCIe x16 slot; 32 GB DDR5 (2 × 16 GB); 2 TB NVMe SSD; 650 W Gold PSU; airflow case.
- What it runs
- Dense models up to about 14B at 4-bit with comfortable context; MoE models like gpt-oss-20b (12.1 GB file).
- Formula prediction
- Qwen3.5-9B Q4_K_M: 448 ÷ 5.63 ≈ 80 tokens/s ceiling. Gemma 4 12B Q4_K_M: 448 ÷ 7.12 ≈ 63 tokens/s.
- Limitations
- 16 GB fills quickly. A 24B model at 4-bit (14.3 GB) fits only with a short context; 27B and up spill into system RAM and slow sharply.
Build B: the 32 GB workhorse
about CA$3,800 to CA$5,000 (estimate)
- Graphics card
- AMD Radeon AI PRO R9700 32 GB (640 GB/s)
- Other parts
- 8-core desktop CPU; B- or X-series motherboard; 64 GB DDR5 (2 × 32 GB); 2 TB NVMe SSD; 850 W Gold PSU; airflow case.
- What it runs
- Dense 24B to 31B models at 4-bit with long context; MoE models up to about 35B total.
- Formula prediction
- Qwen3.6-27B Q4_K_M: 640 ÷ 16.82 ≈ 38 tokens/s ceiling. Mistral Small 3.2 24B: 640 ÷ 14.33 ≈ 45 tokens/s.
- Limitations
- AMD software support is good in llama.cpp (Vulkan or ROCm) but narrower than NVIDIA’s elsewhere. 70B dense models do not fit.
Build C: the fast large-model rig
about CA$8,000 and up (estimate; graphics card street prices volatile)
- Graphics card
- NVIDIA GeForce RTX 5090 32 GB (1,792 GB/s)
- Other parts
- 12- to 16-core desktop CPU; X-series motherboard; 128 GB DDR5 (2 × 64 GB) for MoE offload; 4 TB NVMe SSD; 1,000 W Gold PSU (NVIDIA’s stated requirement); large airflow case.
- What it runs
- Any dense model up to about 31B at 4-bit very fast; gpt-oss-120b (63.4 GB) by keeping attention on the card and experts in system RAM.
- Formula prediction
- Qwen3.6-27B Q4_K_M: 1,792 ÷ 16.82 ≈ 107 tokens/s ceiling. gpt-oss-120b with experts in dual-channel DDR5-6000 (about 96 GB/s, derived): 96 ÷ 2.77 ≈ 35 tokens/s ceiling.
- Limitations
- 575 W card: power, heat, and noise are real. The RAM, not the card, sets the speed once a model spills over.
As of 16 October 2026. Graphics card specifications from NVIDIA and AMD; model sizes from the GGUF listings in lesson 9.7. Budgets are planning estimates: neither NVIDIA nor AMD publishes Canadian prices, and street prices for high-end cards moved sharply in 2026. Price the parts on the day you buy.
The prediction for gpt-oss-120b in Build C uses the MoE rule from lesson 9.2: its 63.39 GB file holds 117 billion parameters, of which 5.1 billion are active, so about 2.77 GB is read per token. [source] With the experts kept in system RAM (llama.cpp's --cpu-moe option), RAM bandwidth sets the pace. [source]
Compatibility checks before you buy
- Power. The PSU must meet the card maker's required system power: 600 W for the RTX 5060 Ti 16 GB, 750 W for the Radeon AI PRO R9700, 1,000 W for the RTX 5090. [source] [source] [source]
- Power connector. Check the card's connector: recent NVIDIA cards accept a 12V-2x6 (PCIe Gen 5) cable or an adapter from 8-pin cables. A PSU with a native 12V-2x6 cable is the tidier choice.
- Size. Compare the card's length with the case's maximum graphics card length, and the CPU cooler's height with the case's cooler clearance.
- Socket and memory. The CPU and motherboard must share a socket; the motherboard decides the RAM type (DDR5) and how many sticks.
- Slots. One full-length PCIe x16 slot for the card; for two cards, check spacing and that the second slot is wired for enough lanes.
Assembly, step by step
Work on a clear table, touch a metal part of the case to discharge static, and keep screws in a dish. The diagram below steps through the build; tap any part name for its definition.
- Lay the motherboard on its box. Lift the CPU socket lever and open the retention frame.
- Align the triangle on the CPU corner with the triangle on the socket; lower it in without pressing. Close the frame and lower the lever.
- Insert the RAM sticks in the slots the motherboard manual marks for two sticks (usually the second and fourth), pressing until both latches click.
- Fit the M.2 SSD into its slot at an angle, press flat, and secure it.
- Apply thermal paste if the cooler has none pre-applied, and mount the CPU cooler; plug its fan into the CPU fan header.
- Fit the motherboard into the case on its standoffs and screw it down.
- Mount the power supply, fan facing the case vent.
- Remove the slot covers and seat the graphics card in the top PCIe x16 slot until the latch clicks; screw its bracket in.
- Connect power: the 24-pin to the motherboard, the 8-pin CPU power at the top left, and the card's power cable, pushed fully home.
- Connect the front-panel switch and USB headers, close the case, and connect a monitor to the graphics card, not the motherboard.
First boot, BIOS, and drivers
On first power-on, open the BIOS (usually by pressing Delete or F2 at start-up). Check that the CPU, all RAM, and the SSD are detected. Enable the RAM's memory profile (XMP on Intel boards, EXPO on AMD), otherwise DDR5 runs at a slower default speed and offloaded layers run slower too. Save and restart.
Install the operating system, then the graphics driver from the card maker. Ollama, for example, requires NVIDIA driver 551.61 or newer on Windows, and an AMD ROCm v7 or Vulkan-capable Radeon driver for AMD cards. [source] Then go to lesson 9.8 for the software.
Self-check
- Which part do you choose first, and by which two numbers?
- Where does the PSU wattage requirement come from?
- What setting do you enable in the BIOS for DDR5, and why does it matter for local AI?
- Build A predicts about 80 tokens per second for Qwen3.5-9B. Show the arithmetic.
Check your understanding
- The graphics card, by memory capacity first and bandwidth second.
- The card maker's specification page ("required system power" or "minimum PSU recommendation").
- The memory profile (XMP or EXPO). Without it, RAM runs slower, which slows anything offloaded to system memory.
- 448 GB/s ÷ 5.63 GB ≈ 80 tokens per second, as a ceiling.