GPUs Are Computers Now
They still draw frames. The queues are for matrix math.
[ essay ]
GPUs are computers now. They still draw frames. The money and the queues are for training and inference: thousands of cores running the same math, memory bandwidth as the scarce resource, CUDA as a dialect that escaped graphics.
I do not work at NVIDIA. I use GPU-backed models the way I use any other computer I do not own: an API, a cloud instance, sometimes a local run on Fedora. The mental model that fails is “graphics card.” A datacenter GPU is a poor fit for a display and a good dense computer. The driver stack, the HBM, the interconnect: those are OS concerns. Gamers noticed when a GeForce became an inference appliance. Writers noticed when “the model” meant a rack, not a paragraph.
If you are capacity planning, ask which computer you are buying. The one that paints pixels, or the one that multiplies matrices. They share an ancestor. They do not share a job. Budget the job. A laptop GPU that can preview a theme is not the same class of machine as the accelerator your vendor is rationing. Calling both “a GPU” is how you under-spec the work and over-spec the game.
The graphics heritage is real. It is no longer the product. Treat the card as a computer with a weird instruction mix, then pick the one whose memory and interconnect match the model, not the one whose RGB matches the case.
— JV · Dark Heart Labs.