“Accelerating GPU software development”
Nvidia recently announced the availability of CUDA 7 Release Candidate (RC). The production version should follow soon. CUDA 7 continues the long evolution of the CUDA language and products for Nvidia Tesla GPU’s. CUDA 7 is available in the current version of CUDA Toolkit. You’ll have to be a CUDA Registered Developer to get the RC version, but of course the production version will be available to the general public.
CUDA 7 includes many new capabilities that should improve the development and performance of GPU-enabled applications.
- C++11 Support. C++11 was released a few years ago (2011). It includes a long list of new language extensions, Standard Template Library (STL) enhancements, and language usability improvements.
- Thrust v.1.8. Substantial performance improvements in various Thrust algorithms, such as sort, merge, scan and others. Speedups ranging from 15% – 300%.
- cuSOLVER. A new numerical library for sparse and dense linear solvers.
- cuFFT. Up to 3X performance improvements for Fast Fourier Transform routines.
- Runtime Compilation Library. Permits highly-tuned, data-dependent runtime compilation of code.
- GPU Core Dumps. Improves code debugging by providing GPU core dumps along with CPU core dumps.
- Memcheck Tools. Minimize software defects by tracking and reporting uninitialized global memory.
- IBM Power8. Supports the new IBM Power8 CPU architecture.
We’ll be applying CUDA 7 in our ongoing development efforts to GPU-enable Next Generation Sequencing data analysis pipelines. Stay tuned for updates on this endeavor.