Cryptographic doesn't mean slow. If you generate in to a large enough buffer (1-2kb is good), Chacha8 is anywhere from ~1.5 cycles/byte (SSSE-3, Wolfdale) to ~0.5 cycles/byte (AVX2, Haswell). Even unoptimized implementations are still fairly efficient at ~5 cycles/byte.

If you don't care about the cryptographic properties, it could be made even faster by dropping to 6 rounds and tweaking the output step.