GPU Rush
Route glowing tokens down conveyor belts into hungry GPUs, keep the context window from overflowing, and upgrade your rig from a potato laptop to a humming datacenter.
Runs 100% in your browser — check the network tabUpgrades
Stats
Under the Hood
Okay, confession: everything in this game is a real thing, just cartoonified. A token really is the small chunk of text an LLM reads and writes one at a time — words, word-pieces, even punctuation each become their own glowing blob. A GPU really is the chip that does the actual math (matrix multiplications, mostly) to turn a token into a prediction — the “crunch” animation is doing you a favor by not showing you a few billion multiplications. And the context windowreally is the hard limit on how much text a model can hold in its “working memory” at once — feed it more than that and, yes, real systems really do throw an error that amounts to “context window exceeded.”
The one honest simplification: FLOPs (floating-point operations) are a real unit of GPU work, but in this game they’re also your currency, which real GPUs do not let you spend on cosmetics.
Go deeper
“More GPUs” and “GPU Clock” map to two real levers datacenters actually pull: parallelism (more chips working at once) and raw clock speed per chip. Real training runs use both — thousands of GPUs, each one also as fast as physics allows.
The prestige mechanic (“Train a Model”) is a wink at how real model development works: you don’t train forever — you ship a version, learn from it, and the nextarchitecture (v2, v3…) starts from scratch but is fundamentally better per-token than the last one, the same way your permanent multipliers make every future run faster.