Performance

pixi run bench produces the figures, in its own environment with PyObjC for the MPS baseline.

GPU

Apple M4 Pro, MLX 0.32.3, one session. A run issues ten calls back to back and evaluates them together. The table gives the best of five runs after a warm-up, in GiB/s of output. mx.random uses its default threefry key. The baseline is MPSMatrixRandomPhilox, Philox4x32-10 from Metal Performance Shaders, timed the same way with ten fills in one command buffer. MPS has no bounded integers and no exponentials, so those rows compare with its uint32 words or float32 uniforms, marked *.

draw	log2 n	tandem_mlx	MPS Philox	mx.random
uniform float32	24	152	162	20
bits uint32	24	165	165	-
randint int32 in [0, 1000)	24	173	166*	12
normal float32	24	170	105	14
normal float32, odd start	24	152	105	-
exponential float32	24	174	163*	-
uniform float32	26	173	145	20
bits uint32	26	164	146	-
randint int32 in [0, 1000)	26	145	145*	11
normal float32	26	159	108	15
normal float32, odd start	26	159	109	-
exponential float32	26	178	164*	-

Tandem and MPS Philox both run near the GPU's integer throughput on the plain draws. The MPS normals run at two thirds of its uniform rate, while the Tandem Box-Muller pass adds little to the fill.