What you actually need.
ART/GTO Solver runs DCFR across a multi-core worker pool on your CPU, and can optionally hand the solve to your graphics card. Real requirements scale with library size. The Viewer runs anywhere a modern browser opens.
Windows machine
Windows 10 or 11, 64-bit. macOS and Linux are not supported. The engine runs on the CPU by default and needs no graphics card: add cores, get faster solves.
Or let the graphics card solve
Since 0.9.35 the Solver can run its iterations and exploitability checks on the graphics card. Set Settings → Solver → Solve on to GPU, or to Auto to use the card only when it helps. It works through the Vulkan or Direct3D 12 graphics driver you already have, with no CUDA and nothing extra to install. Since 0.9.37 a spot too big for the card still solves on it: the solve data stays on the card in a compact 16-bit form, or streams from system RAM. The PC's RAM still has to hold the spot. The CPU stays the default.
One measured example: a full flop solve at 250 iterations, the app's default, took about 260 seconds on the CPU (Ryzen 7 5700X, 8 cores) and about 22 seconds on the GPU (GeForce RTX 4070 SUPER), with ART/GTO 0.9.36. That is one machine and one spot; your times depend on your hardware. Tested on one NVIDIA card so far; AMD and Intel cards are untested.
Disk space scales with depth × coverage
Two dials drive your storage budget: how deep you solve and how many spots you cover. Per-spot sizes range 115–230 MB, and a full library sits in the 0.9–1.3 TB band. A single 2 TB external USB-C drive holds the whole thing. These figures are for the .artgto format as written by 0.9.20 and later. A library exported before 0.9.20 is roughly 3.5 times larger until you re-export it, which --reexport-artgto does without re-solving.
Anything that runs a browser
The Viewer is a self-contained web app served from a one-line local HTTP server. Open it in Chrome, Safari, Firefox, or Edge. It reads .artgto files directly off disk and runs entirely offline once the server is started.