Precision maths libraries for HPC servers

ARM maths libraries take advantage of each silicon partner’s specific micro-architecture innovations and features within their SoCs based on the ARMv8-A architecture to ensure peak system performance. The HPC community had the opportunity to see the first public demonstrations of ARM Performance Libraries at the recent SC15 conference in Austin, Texas.

“ARM Performance Libraries are another key milestone in the advancement of the ARM server and HPC ecosystem,” said Hobson Bullman, General Manager, developer systems group, ARM. “The HPC community are early adopters of ARM-based servers and the introduction of optimised math routines build a foundation for enabling scientific computing on 64-bit ARM-based computer platforms. ARM Performance Libraries hit the mark for both software performance and a consistent library interface to enable ease of software development and portability to ARMv8-A server platforms.”

ARM Performance Libraries optimised for partner silicon

ARM enables load optimised silicon and server platforms that are expanding what is possible for the HPC community. To ensure end customers see compelling results from differentiation enabled by the ARMv8-A architecture, ARM works with its silicon partners to enhance performance tuning for any 64-bit ARM SoC. This approach takes advantage of innovations and design choices such as memory hierarchy and pipeline configuration which can directly influence the throughput of computational operations.

Numerical Algorithms Group (NAG) at the core of ARM Performance Libraries

The NAG Library and associated validation suite is the world’s largest commercially available collection of numerical and statistical algorithms. The NAG Library represents a tried and tested baseline from which to build variants designed for the ARMv8-A architecture. Building on NAG’s industry tested BLAS, LAPACK and FFT maths routines enables ARM to bring mature, accurate and validated core libraries to market quickly. ARM Performance Libraries are multi-threaded, optimised for Advanced SIMD and built using the industry’s latest compilers, enabling peak throughput on 64-bit ARM SoC designs.

“Utilising NAG’s Linear Algebra and FFT algorithms at its core, ARM Performance Libraries are the perfect building block for higher level scientific computing applications,” said Mike Dewar, CTO, NAG. “The hardware specific performance tuning enabled by ARM Performance Libraries, combined with NAG’s stringent requirements for robustness and mathematical correctness, will yield the fastest and most accurate computational mathematics routines for the ARM ecosystem.”

“Cavium is pleased to partner with ARM and NAG to deliver the optimised libraries that are required for the most demanding HPC and Big Data Analytics software applications,” said Larry Wikelius, director, ecosystems and partner enablement, Cavium. “ThunderX, Cavium’s 48-core ARMv8-A-based Workload Optimized Processor with dual-socket support is the ideal platform for delivering best in class performance for high performance computing applications utilising the ARM Performance Library. Cavium and ARM continue to aggressively drive the ARMv8-A server ecosystem with industry leading system and software partners that are delivering key solutions to the HPC, enterprise and hyper-scale server market.”

Enabling open source for ARM HPC

To accelerate porting of software to platforms based on the ARMv8-A architecture, including 64-bit ARM Cortex processor-based servers, ARM Performance Libraries offer tested binary distributions of popular open source applications for HPC including ATLAS, OpenMPI, NumPy and TAU. All changes required to port these open source applications to ARM will be fed back to the open source repositories to enable HPC community development and collaboration.

Commercially licensed

ARM Performance Libraries are available for commercial licensing and supported by ARM, providing a non-restrictive, fully supported, royalty-free and maintained development platform. The pre-built, open source packages will be delivered under their respective license.