The scale of modern compute.

A PlayStation 5 delivers about 10 teraflops (TFLOPS) of performance, which was top-of-the-line in terms of console gaming only a few years ago.

The successor to the PS5, the PS5 Pro, was released in 2024 with ~16.7 TFLOPS of performance and the same 16 GB of unified shared memory (shared between the CPU and GPU, similar architecturally to Apple's M-series chips).

An AI-oriented workstation has slightly different requirements than a typical workstation. The latest line of cards from NVIDIA, the RTX 6000 Blackwell, delivers ~125 teraflops of standard compute and includes dedicated tensor units that assist with model training.

Daisy-chaining cards is where things scale quickly in both power and pricing. When paired, they offer well over 200 teraflops of compute and almost 100 gigabytes of combined memory in a single workstation.

Cloud computing has its advantages, but it can become expensive, regardless of the provider. Now, tasks that once required racks of servers can run locally if the hardware is built for it.

Moore's Law reminds us that over three to five years, it may not seem like much progress is being made, but over a decade or two, we see the leaps that technology continues to make.

• TFLOPS is a valuable measurement, but it doesn't provide a complete description of what's required for an AI-oriented workstation.