LOCAL AI / YOUR MODELS. YOUR MACHINE.
Fit the model.
Tame the watts.
Start with the job: fix a slow run, choose a GPU or build a box with room to grow. Follow the memory, power and cooling decisions in order, then put the machine to work.
Find your next step ↓Hardware picks use maker and project sources; our own measurements are marked in each guide.
LOCAL AI / YOUR NEXT STEP
Diagnose throttling
Read the reason before you change the limit. A power cap and a hot GPU need different next steps.
SW Power Cap or thermal throttling?
Start here when NVIDIA reports a clock limit. Separate power, heat and external braking.
Set a power limit that earns its keep
Follow the nvidia-smi steps and our recorded RTX 3060 Ti cap sweep.
LOCAL AI / YOUR NEXT STEP
Pick a GPU—or a different kind of box
Choose the machine around the model. Then choose a card that fits the case.
GPU desktop or unified-memory mini PC?
The whole-system choice: model size, memory capacity and software route.
Full-size GPU picks by VRAM
Already building a desktop? Compare the cards and the models they suit.
RTX 2000 Ada vs RTX A2000
For a slim workstation: low-profile brackets, slot power and cooler clearance.
LOCAL AI / YOUR NEXT STEP
Build the box
Give every card a slot, every power plug a cable and the host a way back in.
Inside the recorded two-GPU B450 build
Read the parts, slot record and measured Ollama workloads together.
Make room for two GPUs
Plan card thickness, board spacing and the air gap before choosing the case.
Give a Proxmox VM its GPU
Work through ownership, IOMMU groups, recovery and guest restart checks.
LOCAL AI / YOUR NEXT STEP
Run the models
Get the model resident, set the context you need, then measure the wait that matters.
Context, VRAM and parallel requests
Understand what changes when the conversation or request queue grows.
Tokens per second vs response time
Separate model loading, prompt processing and generation from the first useful answer.
RTX 3060 Ti vs RTX 2060: measured workloads
Compare fully resident runs separately from runs that include CPU offload.
When a matched 96GB kit helps
Choose system-memory capacity for CPU inference or offload, with platform checks.
LOCAL AI / YOUR NEXT STEP
Power and cooling
Size the supply for the hardware. Tune the running workload after that.
Choose a PSU for two GPUs
Match output capacity and independent cable runs to the complete build.
Cool the box without the racket
Choose fans and CPU coolers by mounting position, clearance and control.
We earn a commission if you buy through our links — it doesn’t change our picks. How we make money → As an Amazon Associate I earn from qualifying purchases.
SUPPORT THE BUILD
Give the GPU a good home.
Choose these parts for the job they fit. The full guides cover dimensions, power and cabling.
- For a PWM fan positionCheck price on Amazon →
Noctua NF-A12x25 G2 PWM, brown
Match the case mounting position and controller using the quiet-cooling guide. Follow the full buying guide →
- For a build needing a third PCIe cableCheck price on Amazon →
Seasonic FOCUS GX-850 ATX 3.1 (v4)
Map the whole load and separate GPU cable runs using the two-GPU PSU guide. Follow the full buying guide →