PC Gaming

NVIDIA's RTX series of cards have two fixed-function blocks that sit beside the regular CUDA/shader cores.

They have RT Cores which are optimized to accelerate the Bounding-volume-hierachy (BVH) traversal and ray/triangle intersection tests, speeding up raytracing operations.

There are also Tensor Cores which are NVIDIA's "AI" cores, they're optimized for mixed-precision matrix multiplication. DLSS 3 uses a Convolutional Neural Network (CNN) for upscaling and that is, essentially, a bunch of matrix multiplications.

These offload some computation onto dedicated hardware so the CUDA cores that handle the bulk shading/rasterizing are not tied up with these calculations resulting in lower time to render a frame which equates to higher FPS.

AMD cards, in the RDNA2/3 chips have Ray Accelerators, which accelerate the ray/triangle tests but the bulk of the RT load (BVH, shading and denoising) are ran on the regular shader cores. They've just announced (this month) that they're adding 'Radiance Cores' in future hardware, which will handle all raytracing functions like the RT Cores.

AMD doesn't have an equivalent of a Tensor Core, FSR is done in software on the standard shader compute units.

So on NVIDIA cards, DLSS upscaling is 'free' in the sense that it doesn't take time away from the shader cores and RT is accelerated similarly.

This is a good video explaining how Raytracing works if some of the terms are strange to you: https://www.youtube.com/watch?v=gsZiJeaMO48

As an aside, this video is from the 3Blue1Brown 'Summer of Math Exposition' video collection where every year there is a contest for who can make the best and most interesting math explainer videos and this video is one of the winners of the 1st year's contest, the playlists on are 3blue1brown's YT. 3b1b is great all around, if you're into that kind of thing.