If I have to chose a 64 bit preimage resistant hash function; will there be any difference in security between
How long will it take an attacker to find a 64 bit preimage (e.g., minutes, weeks or years)?
SipHash doesn't claim to be a secure hash function. Only a secure MAC. So if you try to use it as a hash function, with a constant, public key, you are on your own.
SHA-512/64 should be a "secure" 64-bit hash, which is of course not enough for a truly secure hash, since it only has 32-bit collision resistance. However, since you only desire preimage resistance, it may have its uses.
How long will it take an attacker to find a 64 bit preimage (e.g., minutes, weeks or years)?
That depends on the attacker's computational power, of course. With a typical desktop CPU you can calculate around $2^{20}$ to $2^{25}$ SHA-512 hashes per second. So a brute force preimage search would take maybe $2^{40}$ seconds, or tens of thousands of years. GPUs could reduce that by a large factor, but it would still likely be centuries.
However, the whole bitcoin userbase is currently (as of early 2016) hashing over $2^{60}$ hashes per second, and even taking into account a small additional factor from SHA-512 vs. SHA-256, you'd be inverting the hash in less than a minute at that rate.
If you want something that takes longer to invert, you should use a password-hashing function, like scrypt with sufficiently large parameters.
A quick resarch showed that there are no (good) attacks on Siphash. For SHA-512 there are defintely no known attacks. The first 64 bits of SHA-512 should have the same security guarantees as full SHA-512 has.
So breaking any of the two comes down to how fast they are.
SHA-512 is slower, in particular it achieves 192.5 cycles / byte in a 64-bit C implementation on an Intel Core 2 Duo with 10 byte messages (enough to find preimages for 64-bit truncation). This means you'd need 1925 cycles to test any potential pre-image. Now let's quickly multiply this by the number of tries you have to do: $1925\times 2^{64} \approx 2^{75}$, meaning you need to invest $\approx 2^{75}$ cycles to find a pre-image for SHA-512. Now let's assume you run a 2GHz CPU, this means you'd need $2^{75} / 2^{31}\approx 2^{44}$ seconds resulting in 557k years a single 2GHz Intel Core 2 Duo CPU would need. Now assume that you have a supercomputer with 1.5 million cores, meaning you'd need $\approx 0.33$ years to find a pre-image with this setup.
Now consider that Intel Core 2 Duos are rather old and SHA-512 can be significantly faster with newer CPUs and that this was a C implementation, whereas an optimized assembler implementation also is somewhat faster. And now also consider that SipHash is faster in usage than SHA-512 meaning that you need even less time.
Concerning whether there will be any difference in security:
No, there shouldn't be any difference in security other than implied by the speed-up when using SipHash.
One needs a few months at maximum if there are enough ressources to find a 64-bit pre-image.
