Buying Guides

Choosing a GPU for running LLMs locally

VRAM is the number that decides whether a model runs at all. As a rough rule of thumb for 4-bit quantized models: 8GB comfortably fits 7B-parameter models, 16GB opens up 13B–20B models with room for context, and 24GB is where 30B-class models and longer context windows become practical without offloading to system RAM. Offloading works, but it's dramatically slower — often the difference between usable and unusable for interactive use.

Compute generation matters less than capacity here. An older card with more VRAM will often run a bigger model that a newer, faster card with less VRAM simply cannot load. Sort our GPU table by VRAM, not by price, if local AI is the goal.

Choosing a GPU for gaming

Here the calculus flips: VRAM matters far less below 4K resolution, and compute (shader count, clock speed, generation) drives frame rates directly. 8–12GB is enough for 1080p and most 1440p gaming today; 16GB+ mainly future-proofs against upcoming titles with larger texture packs, or is needed for 4K. Check the card's TDP against your power supply's headroom before buying — a card that draws more than your PSU can reliably deliver is a return, not an upgrade.

Choosing a CPU to pair with a GPU

For AI training and inference, the CPU's job is feeding the GPU — data loading, tokenization, preprocessing — not doing the heavy compute itself. A mid-range 6–8 core CPU rarely bottlenecks a single consumer GPU. Where core count matters more is multi-GPU setups, heavy data pipelines, or running several models concurrently. Don't under-spend on the GPU to afford a CPU upgrade the workload won't use.

Choosing an edge AI accelerator

Accelerators like the Google Coral or Hailo-8 aren't GPU substitutes — they're for running an already-trained model efficiently on low-power hardware: a Raspberry Pi doing on-device object detection, for instance. TOPS is the headline spec, but form factor matters just as much in practice: USB accelerators are the easiest to add to an existing build; M.2 modules need a free slot and usually deliver better sustained performance without a USB bottleneck.

New vs. used, in practice

• Used GPUs that were used for gaming rather than 24/7 mining typically have an easier life — ask the seller if they know the card's history.
• Check the return policy on any used listing before buying — a short or no return window on a used GPU is a real risk, since early failure is the main way these go wrong.
• New hardware carries a manufacturer warranty (typically 2–3 years on GPUs) at a price premium over used — worth it if you plan to run the part hard or can't tolerate downtime.

Once you know which spec matters for your use case, head back to the comparison table and sort by it.