CUDA 7.5 Up To RC Status With New 16-bit FP Data Format
NVIDIA announced from the International Machine Learning Society (ICML) conference the release candidate for CUDA 7.5.
CUDA 7.5 brings a 16-bit floating point data format (FP16) for storing more data in GPU memory while reducing memory bandwidth requirements, new cuSPARSE GEMVI routines, and intruction-level profiling. The low-level profiling is for helping to find performance bottlenecks within the CUDA code.
Those wanting to grab the CUDA 7.5 Release Candidate for supported platforms can find it over on the NVIDIA developer site.
NVIDIA also announced cuDNN 3 (their CUDA Deep Neural Network library) and DIGITS 2 (another deep learning software piece) at this week's conference.
CUDA 7.5 brings a 16-bit floating point data format (FP16) for storing more data in GPU memory while reducing memory bandwidth requirements, new cuSPARSE GEMVI routines, and intruction-level profiling. The low-level profiling is for helping to find performance bottlenecks within the CUDA code.
Those wanting to grab the CUDA 7.5 Release Candidate for supported platforms can find it over on the NVIDIA developer site.
NVIDIA also announced cuDNN 3 (their CUDA Deep Neural Network library) and DIGITS 2 (another deep learning software piece) at this week's conference.
Add A Comment