AVX threefry kernels for jax (WIP) iirc roughly 1.7 times faster than the autovectorised ones in in the tf part of jaxlib (which already works amazingly well) achieved by upping ILP through manual unrolling