Hi all,
To the best of my understanding, the theoretical timings for FNO are calculated using the formula (total virtuals/active virtuals)^5. For example, if we have 10 total virtual orbitals and 1 of them is frozen, the theoretical speedup we get would be (10/9)^5, or 1.69x. However, it is possible for the actual speedup to exceed the theoretical speedup. like theoretcially we are getting 1.69x of speed but actually we are getting 3.5x of speed
Hi, could you please elaborate your query with the raw timings data and the Q-Chem input you used for getting it?
Performing timings on toy systems with 10 orbitals most likely will result in unreliable timings because it is simply way to small. In this regime, simple data movement can take more time than the actual computation. I would recommend measuring timings on something much more substantial with several hundred orbitals.
It also probably shouldn’t be a 5th-order speedup. It’s not clear what method is intended here, perhaps it’s MP2 due to the 5th-order scaling assumption. In that case, the rate-limiting step scales like o^2 * v^3 so I would expect cubic speedups as virtual orbitals are frozen.