Power Systems

How Power Systems and OpenPOWER enable acceleration

Share this post:

This is the era of accelerated computing.

Today, modern data-centric applications are driving a need for much higher I/O throughput. These performance levels are magnifying the shortcomings of traditional I/O architectures. The only way to turn today’s masses of data into competitive advantage is with an accelerated IT infrastructure built for continued advancement in advanced analytics, machine learning, deep learning, AI and cognitive computing.

Recognizing the need for acceleration and greater processing performance, technology companies are making acquisitions in this space to remain competitive. However, acceleration is not new to IBM. Long ago, we recognized the limitations of Moore’s Law and the role acceleration would contribute in advancing processing capability.

Across eight generations of Power microprocessors and IBM Power Systems™ servers, IBM has always maintained a focus on system value and balanced system design, including high performance scalable I/O and memory subsystems. Others have recognized IBM’s impressive I/O subsystem. See what Google said about POWER at OpenPOWER Summit 2016.

What makes IBM well-positioned for the era of accelerated computing? It’s our collaboration with industry-leading partners to radically redesign the platform at the chip and system level to take advantage of a wide range of accelerators to achieve greater levels of performance than what is available on traditional x86-based servers. These interconnected innovations are collectively known as POWERAccel, and will continue to advance the platform in these application areas:

  • Traditional I/O subsystem. Supports PCIe and NVMe devices with very high sustained I/O throughput.
  • CAPI architecture. Integrates accelerators more tightly into the system so they run natively as part of the application. This provides higher performance and much lower overhead compared to attaching accelerators through a traditional I/O subsystem.
  • A faster, more resilient way to attach memory to a processor, with flexibility to support alternative memory technologies in addition to DDR DRAM.
  • NVIDIA® NVLink™. A high-bandwidth, energy-efficient interconnect that enables ultra-fast communication between the CPU and GPU and between GPUs. This technology enables dramatic speed-ups in data movement and application performance.

We haven’t stopped there. In response to continued rapid growth in demand for bandwidth, acceleration and efficiency, IBM has built on these accomplishments to deliver additional capabilities in POWER9™:

  • PCIe Gen 4. POWER9 is the industry’s first microprocessor to support PCI Express 4.0 to deliver twice the bandwidth of PCIe Gen 3.
  • CAPI 2. Extended the CAPI architecture to operate at PCIe Gen 4 speed.
  • Supported on some models of the POWER9 processor.
  • NVIDIA® NVLink2™. Increased performance and capability.
  • OpenCAPI. The next generation of CAPI for the POWER9 processor. Through the foundation of the OpenCAPI consortium, this new, open technology is now available to everyone — bringing even faster performance in our data centers.

POWERAccel and OpenPOWER

IBM, a founding member of the OpenPOWER Foundation, brings best-of-breed technologies to play with open innovation and continues to collaborate and develop technologies with other leading companies to counter the recent slowdown in hardware innovation.

IBM chooses to be open and makes these technologies (including POWERAccel interconnects) available to allow industry partners and competitors to accomplish their own goals. CAPI has been available to all OpenPOWER members to enable innovation with POWER8®, and many IBM and non-IBM solutions have been brought to market. Openness will only continue, and CAPI 2 and OpenCAPI will be open and available to the OpenPOWER community.

The need for efficient hardware has never been greater. Increasingly complex cognitive workloads require increased processing performance. You can address these performance requirements by adding acceleration technologies around the CPU and selecting from best-of-breed options. OpenPOWER servers with the POWERAccel family of interconnects enable accelerators to compute work faster and more efficiently.

Click here to learn more about the POWERAccel family of technologies that differentiate IBM OpenPOWER servers for advanced analytics, machine learning, deep learning, AI and cognitive computing.

IBM Fellow and Chief Engineer of Power Systems

Add Comment
2 Comments

Leave a Reply

Your email address will not be published.Required fields are marked *


Mr. Kiran N. Mehta

How do FPGA-POWER9 speeds compare with those of NVLink2?


AN

The most exact answer to your question is: it depends. On the workload. If your workload is highly parallel without the need for sequential operations, GPUs are better (like video or image processing or fluid dynamics). Otherwise if it requires “more complex” operations than just a whole bunch of summations, then an FPGA does a much better job.

More Power Systems stories

Demystifying AI for the enterprise at IBM Think

Is AI hype, or does it drive real business benefits? What are the main considerations in building an AI strategy that enterprises need to be aware of? These are some of the questions that I had the privilege of asking a panel of AI experts from IBM, NVIDIA and Gara Guru at the IBM Think […]

Continue reading

Choice and flexibility with IBM & SAP

It’s no surprise that SAP HANA is a resource-intensive workload that requires planning across your organization, from DevOps to infrastructure teams. Last week at the IBM Think Conference in Las Vegas, IBM and SAP leaders talked about how our two companies are innovating together to reduce total cost of ownership and improving efficiency for clients […]

Continue reading

Deep learning & HPC applications get even faster

Data is coming at us from every direction, and it’s up to data scientists and IT leaders to make sense of it by cleaning, processing and extracting actionable insights. It can be challenging enough to prepare your data pipeline, and once you complete this task, you don’t want hardware components to limit your ability to […]

Continue reading