The design and technology used in the GPU. Newer architectures generally offer better performance and efficiency. The three main manufacturers are:

NVIDIA: The leader in AI and professional workloads, featuring CUDA cores, Tensor Cores, and technologies like DLSS and Ray Tracing. Offers premium features and software support for AI through CUDA and frameworks like TensorFlow and PyTorch.

AMD: Delivers excellent price-to-performance value, often providing more VRAM at lower price points. Traditionally stronger performance value in gaming and unable to match NVIDIA's AI capabilities, but with recent partnerships (including OpenAI), their AI capabilities are rapidly improving.

Intel: The newest competitor in the discrete GPU market with their Arc series. While still maturing, they offer competitive ray tracing performance and good video encoding. Their XeSS upscaling technology shows promise, and they're making strides in driver optimization.

Dedicated memory for the GPU. More VRAM allows for higher resolutions and better performance in games with high-quality textures. By being able to hold more information at once, VRAM is also a vital component for running AI models locally, as it determines the maximum size of the AI models that can be processed.

The number of processing units in the GPU. NVIDIA calls them CUDA Cores, while AMD uses Stream Processors. More cores generally mean better performance, but architecture matters too. Example: An RTX 3060 has 3,584 CUDA Cores.

The maximum frequency the GPU can reach under load. Higher clock speeds mean better performance, but also more heat and power usage. Example: RTX 3060 has a boost clock of 1.78 GHz.

The width of the memory bus and the speed at which data can be read from or stored to the VRAM. Higher is better for performance.

The maximum amount of wattage required and heat the GPU is expected to produce under load. This affects cooling requirements and power supply needs. Example: RTX 3060 has a TDP of 170W.