This talk will provide an overview of new debugging and profiling features added in the CUDA 8.0 Toolkit.
The new CUDA Toolkit 8 includes support for Pascal GPUs, up to 2TB of Unified Memory and new automated critical path analysis for effortless performance optimization. This is the most powerful and easy version of the CUDA Toolkit to date.
Pascal architecture and CUDA 8 are particularly efficient for deep learning, neural networks and optimization problems. How can we bring these evolution-like approaches to scientific methods to take full advantage of new GPU architectures? According to Darwin, evolution requires long timescales, while massive parallelization can solve the problem of time. We show how parallel by design scientific methods speed up significantly "in-silico evolution" and present two examples: (i) A parallel version of the Single Chain Mean Field theory, which performs a massively parallel evolution of chain-like molecules by chemical sequence search. (ii) A parallel version of the Bond Fluctuation Model lets us dive into the physics of self-assembly found in living cells on the level of a cellular automaton.