8-Bit Fluids

Abstract

Memory capacity and bandwidth are key constraints on GPU fluid solvers at the scales needed for artificial life simulations. This paper presents a GPU-accelerated Stable Fluids solver with an 8-bit block-floating-point (BF8) encoding at memory round-trip points, reducing bandwidth by 1.78x versus fp16 while preserving greater than 12.7 effective bits of velocity precision. A dithered int8 quantization step at the checkpoint tap-off sets the simulation's noise floor, and because BF8's numerical error falls below this, its effect is negligible. Subtractive dither whitens the noise spectrum, and the solver's inherent dissipation attenuates high-frequency components over time. The fused spectral section eliminates one memory round-trip but requires a hand-implemented 1D FFT. BF8 quantization reduces GPU memory footprint and accelerates spatial processing stages; the spectral section incurs modest overhead from a hand-coded 1D FFT, which an optimized implementation would eliminate. On a 256^3 grid (RTX 3090), the fp16 solver achieves 190 timesteps per second. Extended version of a late breaking abstract accepted at ALife 2026.