Rackspace is at SuperComputing 17 today, showing off our progress on Zaius and Barreleye G2, the project we are developing with Google. We’ll be displaying our latest development samples and offering an OpenCAPI demo.
We’re excited to show the progress we’ve made, especially in the accelerated processing eco-system. Consistent with what we last shared, our samples show significant improvement in CPU performance, IO bandwidth and the latest in coherent-accelerated processing (via new OpenCAPI and NVLink interfaces). We are on track with our original performance projections — no small feat in an era of rapidly diminishing improvements in pure performance and performance-per-watt.
At this stage of server development, we are qualifying a variety of high-speed IO adapters. We are in a unique position to qualify the conventional PCIe adapters (both Gen 3 and Gen 4) and high-speed OpenCAPI accelerators at the same time. Those who do server development for a living will tell you that some of the most-costly, most time-consuming work comes in testing and qualifying add-on components.
Whether you want to use, sell and/or bundle new technology with this platform, you will benefit from these efforts.
PCIe, OCP Mezz and OpenCAPI form-factor adapters being qualified with POWER9 Barreleye G2
What is OpenCAPI, and why does it matter?
OpenCAPI is a new, high speed interface for adapters (network, storage) and accelerators (GPU, ASIC, FPGA), offering special features not found elsewhere. At 25 GT/s bandwidth, OpenCAPI / NVLink2.0 is roughly three times faster than PCIe Gen 3 (the version found in almost every other modern server). Across 32 lanes, that’s a total of 200GB/s bi-directional bandwidth.
OpenCAPI latencies are a huge plus as well, and projected to be an order of magnitude lesser than PCIe standard. OpenCAPI devices operate natively within an application’s user space and coherently with POWER9. This allows an accelerator (GPU, ASIC, FPGA) to operate with little or no overhead in the CPU core, and the operating system kernel.
Those who work on systems with accelerators can attest to what an enormous challenge this is, both in the overhead imposed on a system, programming difficulties, and the lack of good solutions to these problems. OpenCAPI provides an elegant solution to these problems, with an open standard, and a great implementation in the POWER9 CPU.
The OpenCAPI accelerator and software ecosystem is growing rapidly, and Zaius/Barreleye G2 is at the center of this ecosystem enablement. With two POWER9 LaGrange CPUs, our server houses 32 OpenCAPI/NVLink lanes and supports up to four OpenCAPI cards or two NVLInk 2.0 GPUs.
This, augmented by the fact that Rackspace and Google are doing the platform development in the open, with design, manufacturing and firmware available via the Open Compute website, makes our system a very appealing development platform for OpenCAPI accelerators. Indeed, many solutions are being built with our system as a development platform. OpenPOWER’s innovative CAPI SNAP platform, which makes it easy for application developers to use FPGA based accelerators, will be also available.
What does mean to end users and customers? Big performance improvements.
First, through POWER9 CPUs, which will be the fastest CPU of its time. This, when combined with the OpenCAPI/NVLink-based accelerators, will provide application-specific speedups, orders of magnitude greater than a CPU can provide. A few of these accelerators are already productized, from companies like Mellanox, Nvidia and Alpha Data, with more coming soon.
With accelerators spanning storage, networking and extremely parallel processing, users of this server can get speedups on a variety of applications, in an array of industries: machine learning, big data platforms and real-time analytics.
DVT Barreleye G2 server with Nvidia Tesla GPUs
Inspired to join us in building applications and solutions based on this new generation of processor and interconnect technology? If you are a developer or solution designer, reach out to us to learn how you can help. We intend to continue support for the ecosystem, continually improving our OCP Specification Package (updated DVT spec here) and offering hardware for development purposes, where possible.
For those who want to participate as consumers, evaluation samples are available upon request.
We look forward to continued work alongside our partners: Google, OpenPOWER, OpenCAPI, Ingrasys, and adapter providers, both OpenCAPI/NVLink and PCI Express.
We aim to announce production-ready systems in the second quarter of 2018. Stay tuned.
Attending SuperComputing 17 in Denver? Visit OpenCAPI booth #1587 for a demo and a conversation with Rackspace and OpenCAPI consortium teams.