For at least some applications, the details of the processor architecture are dominated by how much high-throughput RAM you can throw at the problem, and GPUs a

For at least some applications, the details of the processor architecture are dominated by how much high-throughput RAM you can throw at the problem, and GPUs are by far the cheapest and most accessible way of cramming a bunch of high-throughput RAM into a computer. While it's not exactly a mainstream solution, some people have built AI rigs in 2022/2023 with used Vega cards because they're a cheap way to get HBM.