Discrete GPU (dGPU)
In plain English: A standalone graphics card with its own dedicated processor, cooling fans, and private pool of high-speed VRAM—such as an NVIDIA GeForce RTX 3090, RTX 5090, or AMD Radeon RX 9070 XT. Unlike integrated graphics built directly inside a standard laptop or desktop CPU that have to share slow motherboard RAM, a discrete GPU plugs into a motherboard's PCIe expansion slot (or sits as a separate chip inside a gaming laptop) and handles heavy parallel math completely on its own.
Why it matters for local AI: For local AI, discrete GPUs are the gold standard for raw speed. Because a dGPU pairs thousands of specialized AI compute cores with ultra-wide GDDR6, GDDR6X, or GDDR7 video memory, it reads model weights and generates text dramatically faster than CPU-only setups or most unified-memory machines. A discrete card also keeps your desktop modular—you can swap in a faster card later or drop a second dGPU into a multi-slot motherboard to combine their VRAM for larger models.
Trade-offs & limitations: Memory capacity, power draw, and heat are the main constraints. Consumer discrete cards top out at 24GB to 32GB of VRAM, meaning a single card cannot load massive 70B+ models at standard precision on its own. Flagship dGPUs also demand hefty power supplies and well-ventilated PC cases to keep temperatures and fan noise under control.
Looking for verified hardware rigs?
Check our real-time £/GB VRAM index to find systems optimized for local model inference.