How is GPU different?

[Francis]

I don't think I understood what the preview levels were trying to show me. How will this game diverge from one where you build a CPU? What is different about building a GPU?

[wheatley]

Fair question, the current previews only show the first step: sharing work between processors. That alone doesn’t explain the GPU angle very well.

The early building blocks overlap with a CPU-building game: logic gates, arithmetic, memory, and a small programmable processor. The difference comes later, when the puzzles focus on running the same program across many pieces of data and keeping those processing lanes busy.

For example, later levels introduce groups of lanes executing instructions together. If those lanes take different branches, some have to wait while others run. One puzzle gives you a program that produces the correct answer but misses its deadline; adding another processing group still isn’t enough, so you need to rethink the program.

Memory becomes another constraint: more processing power won’t help much if everything is waiting for data. You work with bandwidth and shared local memory, then balance processing capacity, memory bandwidth, and power as you scale up.

CPUs also use parallelism, so “more cores” isn’t the distinction by itself. It’s those execution and memory tradeoffs that the later game explores. Your feedback makes sense: the previews currently show the transition toward that, but don’t yet demonstrate what makes it interesting

[Francis]

Cool. If I finish the final game, will I know what things from my GPU specs like shaders are?

[wheatley]

Mostly yes, at least for the compute side of a spec sheet.

The "shader cores" (NVIDIA calls them CUDA cores, AMD calls them stream processors) are the processing lanes you build and keep busy in the game. Those lanes are grouped into the larger units on your card's spec sheet, called SMs or Compute Units, and the warp cores you build in the later levels are a small version of those. You'll also learn why memory bandwidth and on-chip shared memory matter as much as the core count, and why power (TDP) puts a ceiling on how far you can scale. One of the later puzzles is a matrix multiply, which is the kind of work tensor cores are built to speed up.

Some things on your spec sheet won't be covered. The game doesn't go into the graphics-specific hardware: texture units, ROPs, rasterization or ray tracing. When you finish, a line like "X shader cores, Y GB/s bandwidth" should mean something concrete to you, but the game won't teach you how a frame gets drawn.