Developing for OpenPOWER and NVIDIA NVLink

On September 8th, 2016 IBM announced a new system, the S822LC for High Performance Computing, code named “Minsky”. Minsky is the culmination of a co-development effort between NVIDIA and IBM to more tightly couple the CPU and GPU through a high bandwidth, low latency interconnect called NVIDIA NVLink™. NVLink delivers greater than 2.5 times more bandwidth than PCIe and allows the four NVIDIA Tesla P100 GPUs access to the massive memory bandwidth and exceptional system I/O of the dual Power8+ CPUs. The combination creates a platform that is capable to deliver extreme performance on both traditional HPC applications and the fast growing High Performance Data Analytics applications that are key to every modern Enterprise datacenter.

Ruchir Puri, an IBM Fellow at IBM’s Thomas J. Watson Research Center shares how they are building large-scale big data systems and delivering real-time solutions, such as using machine learning, POWER CPUs and NVIDIA GPUs to predict drug reactions.

Minsky with NVLink APPLICATION PERFORMANCE PROJECTIONS

The Minsky platform provides optimized system for moving data throughout the system. The Power system is designed for maximum bandwidth between the CPU and memory and I/O subsystems. The addition of NVLink between CPU and GPU deliver a platform that is ideal for application acceleration. As you can see from Diagram 1, Minsky delivers uncompromised data movement between and GPU to main system memory without a bottleneck. The results of this combination can be better realized applications scalability as shown in Diagram 2 for MILC with increasing lattice sizes and near linear scaling to four Tesla P100 GPUs.

If you’d like to try out your code on this platform or want more information on the other resources available to port to GPUs or the Minsky platform look at the following options.

How to Run Code on POWER + GPU Machines

  1. IBM NVIDIA NVLink Acceleration Lab:
    Supports customers looking to port new applications to the platform with support from IBM and NVIDIA experts.
    To apply for the lab please click on the following link and follow theenrollment directions.

  2. Running Applications:
    Getting your application ready to run on POWER is a simple recompile and run. There are more instructions available through our blog post.

  3. Libraries and CUDA:

    CUDA 8.0 is now available for POWER and can be downloaded today.

All libraries that are included in the latest CUDA 7.0 release are ported to POWER and ready for you to use.

For additional information contact:

Julia Levites , NVIDIA