What does a GPU sound like when it lies? Not crashes — those are honest. The interesting failures are the quiet ones. The precision that degrades so slowly you don't notice until the grid dissolves. The buffer that forgets to clear and lets time accumulate in the framebuffer. The integer that wraps silently and tiles the universe.
Menkman catalogued the vernacular of file formats — JPEG's 8x8 block artifacts, PNG's scanline prediction residuals, GIF's dithered palette quantization. Each encoding fails in its own signature way. The error IS the fingerprint of the medium.
We asked the same question of the GPU. Not the file format but the compute substrate. What are the native failure modes of floating-point arithmetic, framebuffer management, integer overflow? And can those failures become a vocabulary — not bugs to fix, but a language to speak in?
Each failure mode above is a live WebGL shader that demonstrates one specific GPU behavior. Move your mouse to control the parameters. The mouse isn't adjusting an effect — it's tuning the conditions under which the GPU reveals itself. Move far enough right in PRECISION LOSS and the float32 mantissa runs out of bits. Adjust the decay rate in BUFFER ECHO and watch temporal smear emerge from a skipped glClear.
These are not simulations of failure. They are failures, running live on your GPU, right now. The only difference between this and a bug report is intention.
The drone is three static oscillators. A1, E2, A2 — an organ in water. Underneath, four failure modes of the digital audio substrate. Toggle each and listen: aliasing, denormals, cent quantization, phase cancellation. These are to the ear what precision loss and buffer echo are to the eye — the medium's fingerprint made audible.
- ← → navigate modes
- 1-4 jump to mode
- P toggle parameters
- L freeze time
- mouse controls failure parameters