Skip to main content
. Author manuscript; available in PMC: 2012 Aug 26.
Published in final edited form as: J Chem Theory Comput. 2011 Aug 26;7(10):3072–3082. doi: 10.1021/ct200360f

Figure 3.

Figure 3

Timing in ms/step for calculations of non-bonded interactions for different GPUs and different options for lookup tables. The rest of the calculations in the GPU presented in this paper were done on the GTX480. The numbers below include force calculations and communication between the GPU and the CPU. Quad-256 is a table with quadratic interpolation of 256 entries that resides in the shared memory. It is the fastest option and retains acceptable accuracy (Table 2). Quad-4096 is a quadratic table with 4096 entries, which is our most accurate option, and only moderately more expensive than the other options. It is too large to fit in the shared memory and resides mostly in the texture memory. Lin-8000 is a linear interpolation scheme that resides mostly in the texture memory and offers no advantage, accuracy or speed-wise, compared to Quad-256. Coalescing the data is critical. If the data of the non-bonded list matrix Mij is not coalesced then running on GTX480 the Quad-4096 option increases the time from 9.9 to 16.5ms