Have we benchmarked this against ChemFP? about gpusimilarity HOT 2 CLOSED

ChemFP is a lot smarter algorithm, where as this relies on it being so cheap to do and just brute forces it. The GPU speed here is calculating the full similarities of ~200 million (max GPU can hold) in 0.4 seconds. When having to do full similarity (which disadvantages the intelligence of ChemFP), it looks like ChemFP can do about ~10 million a second, so this is 50x faster than ChemFP when it needs to do a relatively stupid thing. Limiting to similarity >0.6, it looks like ChemFP can screen about 20 million a second, or 25x slower.

From my point of view, once you're up to 500 million a second throughput, it makes the intelligence of the algorithm dramatically less important. I'll likely do some type of hierarchical improvements to the algorithm at some point here, not for speed but to enable screening billions of molecules when the GPU can only fit 200 million. For the moment though, this seems to be doing good enough for its primary use.

from gpusimilarity.

We can always halve the fingerprint size and store 400 million :)

from gpusimilarity.