If I try to create collisions for MD5, I can make one every 14 seconds (on average) on my PC, using a single core (Core2, 2.4 GHz). This exploits the weaknesses in the internal structure of MD5. If I only try random data and wait for collisions to appear, well, I will wait for quite some time: the first collision is expected after about 264 hashed messages (give me a thousand PC, and I should achieve a collision in about 20 years of full-time computation).
For currently unbroken cryptographic hash functions, there is no known internal weakness (that's what "unbroken" means), so trying random messages is the best known method to create collisions. Chances to get a collision this way are vanishingly small until you hash at least 2n/2 messages, for a hash function with a n-bit output. This means that with any proper hash function with an output of 256 bits or more, the collision rate is, in practical conditions, zero (you will not get any and that's the end of the story).
Wikipedia has some pointers on the subject. See also chapter 9 of the Handbook of Applied Cryptography (page 369).