GTC 2015 The OpenPower Foundation unveiled new hardware – including its first commercially available server – at its first OpenPower Summit on Wednesday, a mini-event nestled within Nvidia's GPU Technology Conference in California.
The foundation has also forged a relationship with the Facebook-led Open Compute Project, which operates in the Intel x86 server space IBM bailed out of.
The OpenPower Foundation came about in 2013 after IBM flicked through ARM's playbook on licensing processor core designs, and decided – with Google, Nvidia, Mellanox, Tyan, and others – to pull the same moves in the high-performance data center world. IBM agreed to liberally license its Power processor architecture so manufacturers can build server boxes running a GNU/Linux stack and other open-source code. Big Blue provides the CPU blueprints, firmware and software, Nvidia provides the NVlink interconnect and GPUs, and Mellanox does the networking.
If you want to give Power8 a spin on OpenPower hardware, IBM has added boxes to its Softlayer cloud.
Tyan server first
This week, Tyan announced its snappily named TYAN TN71-BP012, the first non-IBM-branded, commercially available Power8-powered server built from OpenPower specifications. It has a single-chip Turismo processor, 32 DDR3 ECC RAM slots, room for 14 hot-swap hard drives, 4 PCIe generation-3 slots, and 4 10GbE ports.
The server will power the aforementioned Softlayer bare-metal service, and reach general availability before the third quarter of 2015.
OpenStack, developer kit, and a prototype
Cloud host biz Rackspace has buddied up with the OpenPower gang to offer OpenPower hardware from its cloud, running OpenStack. The foundation's first GPU-accelerated developer system, the Cirrascale RM4950, will ship in volume before the third quarter of 2015: this has a four-core Turismo, and can take up to four Nvidia Tesla GPUs. IBM is working on new servers, with its pals Nv and Mellanox, codenamed Firestone.
Open Compute, come on down
OpenPower is not the only consortium touting liberally licensed hardware designs for the data center: Facebook's Open Compute Project (OCP) is at it as well. You might say it is a rival effort, but in any case – the OCP has joined OpenPower's new advisory group, along with the Linux Foundation, and the China POWER Technology Alliance (CPTA). These advisers "will provide guidance to the OpenPOWER Board of Directors and serve as a forum for support and collaboration between communities with open approaches to infrastructure and software development," we're told.
What the execs say
In a Q&A session with the press on Wednesday, Gordon MacKean, chairman of the OpenPower Foundation and a senior director at Google, tried differentiating OpenPower and Open Compute:
OCP is about managing and creating a cost-effective platform while OpenPower is about leveraging IBM's Power architecture, and the performance side of the equation: the performance per dollar, we focus on that.
Brad McCredie, president of the OpenPower Foundation and an IBM fellow, added:
OCP is now on our advisory board. The OpenPower Foundation has created an advisory group, the first to join was the Linux Foundation, second was Open Compute, and the third member is the China POWER Technology Alliance, which is a 30-member group in China. The group will advise the board of directors of Open Power.
McCredie also insisted OpenPower hardware will not be a poor cousin to Big Blue-branded, Power-powered hardware.
The strategy for IBM in terms of contributing to the OpenPower Foundation is just as you saw for Power8: we participate in OpenPower, continue to introduce products into the IBM product line, and the technology will be immediately available to OpenPower partners. We need to keep that time lag from the introduction of IBM product lines to the OpenPower partners as small as possible.
MacKean told El Reg that "OpenPower will be pushing hard on IBM to get Power9" as soon as possible; the architecture is due to land around 2017, as that's when the US Department of Energy will get a couple of supercomputers packed with the chips.
McCredie also touched on a growing demand for source-code visibility from security-conscious big business, although he might well say that as it's a selling-point for the foundation:
There are some trends in the industry worldwide and one is that we may get to the point where the definition of secure is "open," ie: if I can’t see the source code, I’m not sure I can trust it. Maybe there’s going to be a bigger trend as security concern escalate, as people become more security concious, and they may want more and more open technology, and that an opportunity for OpenPower to take advantage.
One other thing we watch is the lines between hyper-scale and enterprise, and they are getting blurred. One of our goals is to bring accelerated computing to enterprise. One of my demos is DB2, IBM's database, being accelerated using GPUs. HPC is a bellwether for enterprise.
Sumit Gupta, the general manager for Tesla Accelerated Computing at Nvidia, added that GPU-accelerated hardware (such as OpenPower designs) doesn't have to be limited to the supercomputer world:
We [at Nvidia] have looked at the database market for a long time, and there are two critical reasons why GPUs can’t play well in that market: one is the connection between the CPU and the GPU, that was hard to do right, and second was the memory on the GPU. We've solved those problems: we have a very high speed interconnect, NVlink, and will be adding 32GB of RAM to our GPUs, so the market for enterprise will dramatically open up.
Back to McCredie, and he said support for DDR4 RAM may come this year. He also welcomed chip makers coming forward to fabricate their own custom Power8 processors using IBM's licensed blueprints:
That would be a success: the ability to do that is something our partners can do with the IP. That will create multiple independent products, all on a common architecture and software API, so all the software runs as you expect, but allowing the architecture to target multiple markets. If that happens, that’s a success. We could see the first one [a non-IBM OpenPower processor] in two years, that's a good solid design cycle.
We need a range of performance-price options. ®