Choose the model and context you need before choosing the graphics card. Dedicated VRAM determines whether the required weights and conversation memory can stay on the GPU; app, driver, and operating-system support determine whether that capacity is usable. A card's gaming position or advertised AI throughput is not a substitute for those checks.
This guide is for someone considering an upgrade for local text models. If you already own a suitable computer, test a small model on it first. For occasional short prompts, CPU-only use may be sufficient and avoids a purchase you do not need.
Write down the workload
Choose one exact model download, one app, the intended context, and whether you need one request or several at once. Include document search or an agent if those are part of the job. A chat using a small model and a few paragraphs is a different workload from an agent carrying repository files and tool output.
Use the RAM and VRAM calculator with those choices. Compare requirements against dedicated VRAM and system RAM separately. Do not combine them into one interchangeable capacity.
Use memory tiers as starting points
| Dedicated VRAM | Starting investigation |
|---|---|
| 6GB | A small 3B–4B quantized model; larger files can leave little context room |
| 8GB | Selected 7B–8B 4-bit models with modest context |
| 12GB–16GB | Selected 14B-class compressed models, depending on exact file and context |
| 24GB–32GB | Selected 32B-class compressed models; still check conversation memory |
| More than 32GB | Larger models or longer context, based on actual requirements rather than the tier alone |
These are capacity planning examples, not compatibility or speed promises. Some architectures need different memory at the same parameter count. A dense 70B model at exactly four bits needs about 35GB for weights alone before format overhead and conversation memory; 32GB is therefore not sufficient for that full four-bit weight allocation.
Check the exact app and driver path
For Ollama, compare the exact card with its hardware-support documentation and the requirements for your operating system. NVIDIA support includes a compute-capability and driver requirement. AMD's supported ROCm devices differ between Windows and Linux, and a Vulkan path is another configuration with its own driver needs.
For llama.cpp, choose binaries or a build using the intended backend, such as CUDA, HIP, Vulkan, Metal, or SYCL. A CPU build running successfully does not establish GPU acceleration. LM Studio also uses installed model runners; check the required runner and model support in that app.
Do not arbitrarily reduce the permitted model size because a card is AMD. First establish the backend, available memory, and actual placement of the workload. Unsupported acceleration may change the recommendation, but the vendor name does not change a file's size.
Check the rest of the computer before paying
- Confirm the card fits the case and motherboard layout.
- Check the card vendor's power-supply, connector, and clearance requirements for the exact product.
- Allow for cooling, fan noise, and sustained power use.
- Keep enough system RAM for the app, operating system, and any CPU-loaded work.
- Budget disk space for model files and document data.
- For a used card, check the condition, return policy, and continued driver support.
Desktop and laptop products with similar names can have different memory and power limits. Check the specification of the actual machine or board. Do not use the desktop card's specification to describe a laptop GPU.
Compare performance only when the comparison is meaningful
Look for the exact model, compression, context, app version, and hardware in any published test. Prompt processing and answer generation are different parts of a workload. A short decoding result does not predict a long document or agent session.
If you can test the candidate, use one of your harmless real prompts and check both response time and answer quality. In Ollama, ollama ps shows CPU/GPU placement after loading. A useful purchase outcome is the required model and context running with acceptable waits and room for the rest of your work.
Multiple cards introduce software, power, space, and communication requirements. Two cards do not automatically act like one card with their combined VRAM. Plan around the specific model runner's placement rules rather than buying a pair on that assumption.
Use the cost calculator for purchase and running-cost scenarios. Prices and electricity costs vary; enter your actual quote and usage. Lower estimated money cost does not establish equivalent answer quality or privacy.