Packed particle layout using 16-bit int/fp struct fields. Borrowed my old ideas/code from Claybook's clay/fluid simulators. Particle pool VRAM usage: 64MB->32MB Performance: 12% faster sim/sort, 7% faster draw Math on GPU is still pure fp32/int32. Not yet fp16/int16.

Sep 23, 2026 · 1:30 PM UTC

1
2
65
3,389
Sort replies: Relevant Recent Liked
Replying to @SebAaltonen
I learned that the CPU has f32 math acceleration. Is that not the case for GPUs or is sending data to get GPU just such a bottleneck that the smaller the better?
7