How does a graphics card (and a GPU) work?

Brett Bergan, Robotics TechnicianOriginal authorThe GPU is the “brain” of a graphics card. Data is transported from the CPU via the PCIe lanes to the GPU, where parallel processing can take place in the GPU cores. The small slot with eleven contacts on the left consists of 12V and 3.3V rails and some IO controllers. The long slot is the PCIe lanes. Zoom in to see the locations of tracks 0 through 15. The GPU chip on this RTX 2060 has 1,920 CUDA cores arranged in 30 64-bit blocks. Each block of 64 (called a streaming multiprocessor or SM) has its own raytracing core. A block of data is sent to the GPU. This is the map of the polygon vertices required to create a wireframe skeleton of the current screen. The GPU divides the workload and sends it to the GPU shader cores. The textures are then called from the VRAM buffers and placed on the polygon wireframes. Shadows, reflections, and other effects provide more realism and detail. Finally, certain post-processing effects such as anti-aliasing (or RTX card ray-tracing) are performed to further enhance and smooth the image. Each PCIe 3.0 lane can transport an enormous 984.6 MB / s of graphics data. So it’s almost impossible that a GPU can ever use nearly its capacity of all 16 PCIe lanes. Even something like an RTX 2080 Ti would seldom need more than two or possibly four PCIe lanes. To be able to render 144 fps, the GPU must be able to handle all of these processes – wireframe to structured overlay of rendered scene to postprocessing effects – complete in less than 6.9 milliseconds


Leave a Reply

Your email address will not be published. Required fields are marked *