TensorFlow is an open source framework for machine learning developed by Google. It provides a very flexible set of tools to create numerical computations of data flow graphs. TensorFlow is achieving tremendous adoption because its design makes its users more productive.
The TensorFlow enablement opens new options for applying cognitive technologies and cutting-edge machine learning algorithms to your mission-critical enterprise data:
- Train and predict by avoiding the unnecessary copying or ETL of huge datasets outside of your most sensitive security zones.
- Develop algorithms on a laptop and seamlessly scale up to production on LinuxONE or Linux on z System, while working on data in place.
- Leverage all of the innovation occurring in the TensorFlow ecosystem.
Advantages of running TensorFlow on LinuxONE and z Systems
LinuxONE supports up to 10TB of memory, allowing huge neural network models and huge datasets, all managed with ease in a single system in memory. This means no fragments and sharding, no extraneous complexity from parallelism, and just pure scale and power.
IBM LinuxONE and Linux on z Systems users can deploy the latest Machine Learning and Deep Learning research ideas to mission critical data without ever leaving the system. Tap the skills of the latest data science and machine learning talent to extract insights from enterprise and mission-critical production data and transactions.
Learn why IBM LinuxONE is a great platform to jump start next generation applications by reading the white paper.
How TensorFlow Works
The foundation of TensorFlow is the Eigen library Basic Linear Algebra Subprograms (BLAS) math routines. Eigen provides a highly efficient and dynamically adaptive BLAS library implemented as C++ header files. The implementation leverages the system cache hierarchy and architecture-specific arithmetic optimizations, including SIMD vector instructions.
Over the past year, Eigen developer Konstantinos Margaritis added architecture-specific optimizations for IBM z13 SIMD architecture — initially double precision floating point vectors and then special support for single precision floating point operations. Full support was included in the recently released Eigen 3.3.1, leveraging the new SIMD floating point instructions for greater floating point performance.
The performance of Eigen on z13 directly flows to the performance of TensorFlow on LinuxONE and Linux on z. And it’s a hot rod.
Find out more about TensorFlow at https://www.tensorflow.org/