G as summarized in Table 1.On single nodes, GROMACS’ built-in thread-MPI library was applied.GROMACS is often compiled in mixed precision (MP) or in double precision (DP). DP treats all variables with DP accuracy, whereas MP utilizes single precision (SP) for most variables, as for example, the huge arrays containing positions, forces, and velocities, but DP for some critical components like accumulation buffers. It was shown that MP doesn’t deteriorate energy conservation.[7] Since it produces 1.43 much more trajectory inside the very same compute time, it’s in most instances preferable more than DP.[17] Consequently, we utilised MP for the benchmarking. GPU acceleration GROMACS 4.six and later supports CUDA-compatible GPUs with compute capability two.0 or greater. Table three lists a selection of modern GPUs (of which all but the GTX 970 were benchmarked) which includes some relevant technical info. The SP column shows the GPU’s maximum theoretical SP flop rate, calculated in the base clock rate (as reported by NVIDIA’s deviceQuery system) times the number of cores occasions two floating-point operations per core and cycle. GROMACS exclu-sively utilizes SP floating point (and integer) arithmetic on GPUs and may, therefore, only be used in MP mode with GPUs. Note that at comparable theoretical SP flop price the Maxwell GM204 cards yield a greater efficient overall performance than Kepler generation cards resulting from improved instruction scheduling and lowered instruction latencies. As the GROMACS CUDA nonbonded kernels are by design strongly compute-bound,[9] GPU primary memory performance has little impact on their performance. Hence, peak efficiency of the GPU kernels can be estimated and compared within an architectural generation simply from the product of clock price 3 cores. SP throughput is calculated in the base clock price, but the effective efficiency will significantly depend on the actual sustained frequency a card will run at, which could be considerably higher. Benchmarking procedure The benchmarks were run for 20005,000 steps, which translates to a few minutes wall clock runtime for the singlenode benchmarks. Balancing the PubMed ID:http://www.ncbi.nlm.nih.gov/pubmed/20148622 computational load requires mdrun as much as a number of thousand time methods at the starting of a simulation. As throughout that phase the efficiency is neither stable nor optimal, we excluded the first 10000,000 steps from measurements making use of the -resetstep or -resethway command line switches. On nodes without having a GPU, we normally activated DLB, because the positive aspects of a balanced computational load involving CPU cores commonly outweigh the small overhead of performing the balancing (see e.g., Fig. 3, black lines). On GPU nodes, the scenario is just not so clear as a result of competitors among DD and CPU PU load balancing mentioned within the Essential Determinants for GROMACS Efficiency section. We, thus, tested both with and with out DLB in most of the GPU benchmarks. All reported MEM and RIB performances would be the typical of two runs each and every, with normal deviations around the order of a couple of percent (see Fig. 4 for an instance of how the data scatter).Determining the single-node overall performance. We aimed to discover the AM152 site optimal command-line settings for each and every hardware configuration by testing the different parameter combinations as described within the Key Determinants for GROMACS Efficiency section. On person nodes with Nc cores, we tested the following settings making use of thread-MPI ranks:Intel E5680v2 3Intel E5680v2 three two with 23 GTXThe last column shows the speedup in comparison with GCC four.four.7 calculated in the average of the speedups on the.