An IBM project to expand the market for its Power processor is making headway, with new hardware announced Wednesday that aims to challenge Intel’s dominance in the data center.
IBM still has a lot of work to do, but the project it launched two years ago to open up the Power architecture for use by other hardware makers is gaining momentum. The idea is to lower the cost of Power-based systems so they can be sold into hyperscale data centers and high-performance computing environments, areas dominated today by x86 processors.
Tyan, a server manufacturer in Taiwan, will deliver the first commercially available OpenPower server in the second quarter, a two-socket system aimed at hyperscale customers such as Internet service and cloud providers, IBM said.
The first OpenPower system from IBM is also on its way, a product known as Firestone that will target HPC users and is due out later this year. It will be manufactured by Taiwan’s Wistron and sold by IBM, and incorporates GPU technology from Nvidia and interconnects from Mellanox.
They are just two of the announcements being made at the OpenPower Summit in Silicon Valley. IBM says the OpenPower Foundation now has 113 members, up from 30 nine months ago.
IBM launched the effort with Tyan, Nvidia, Mellanox, and Google. Sales of IBM’s traditional, Power-based Unix systems are declining, and through the OpenPower Foundation it hopes vendors will produce a variety of systems that will breath new life into the architecture.
It’s doing it in a couple of ways. First, it provided hardware reference designs for Power servers, along with open source firmware, virtualization and other software, to let third parties like Tyan build their own systems using Power 8 processors.
That’s already a big change for IBM, which historically kept its Power technologies close to its chest. But IBM’s business model isn’t set up to sell low cost systems, and it’s encouraging hardware partners to produce servers in the sub-$6,000 range to reach new customers.
IBM is also licensing the Power architecture itself, in a model similar to that used by ARM, so other companies can build derivatives of the Power chip. That's a longer-term effort, but IBM is also announcing that China's Suzhou PowerCore expects to release the first non-IBM Power server chip in the second quarter. Dubbed CP1, it will go into a server being developed for the local Chinese market.
IBM has big hopes for China, where policies favoring local suppliers are making it harder for outside firms to compete. It hopes the homegrown systems will appeal to big cloud providers like Baidu, Alibaba and Tencent.
“It opens up China for us, because we’re now talking about a local producer instead of a multinational,” said Ken King, IBM manager for OpenPower alliances.
Despite the momentum, some see challenges ahead. Creating new hardware is one piece of the puzzle, but IBM has to convince customers that cost and performance gains from Power will be sufficient to justify introducing a new chip architecture to environments that have standardized on x86.
“So far, this is all very embryonic,” said Nathan Brookwood, principal analyst at Insight64.
IBM says customers will be able to port Linux applications to Power relatively easily, especially now that Power 8 supports the “Little Endian” compute model used by x86 chips. But those applications will still need to be patched and maintained separately from their x86 counterparts, which adds a layer of complexity that hyperscale customers like to avoid, Brookwood noted.
He thinks customers are expressing interest in Power to gain pricing leverage against Intel.
“If you look at what a Tyan or a Rackspace has to do to put together a Power system, it’s not a huge expense, and there are a whole lot of people who buy Intel servers who right now feel Intel has them over a barrel,” Brookwood said. “They would love to leave an OpenPower brochure on the desk when the Intel sales guy comes around, to let them know they’re not the only game in town.”
IBM, obviously, doesn’t see it that way. It says Power can offer significant cost-performance advantages, especially for big, data-intensive applications, thanks to its high thread count, on-chip bandwidth and the large amount of memory it can address.
It has a partial vote of confidence from Google, which said last year it had built its own Power-based server board. It’s still not known whether Google plans to use Power servers in production, though one of its engineers is chair of the OpenPower Foundation.
Rackspace is also a fan. It’s designing a Power 8 server that it will use to offer a new, bare metal compute service running the OpenStack cloud software. Aaron Sullivan, Rackspace senior director and distinguished engineer, showed an early version of its server board, dubbed Barreleye, at the Open Compute Project Summit last week. “This will be much more efficient for our apps than anything we have today,” he said in a talk at the event.
The other new hardware announced Wednesday includes a developer platform from Nvidia, Tyan and Cirrascale, for building GPU-accelerated systems that run big data analytics, deep learning and scientific computing applications. Called the Cirrascale RM4950, IBM says it will ship in volume in the second quarter.
Xilinx and Altera are also on board. They say they’re working with IBM and its partners to allow programmable chips, known as FPGAs, to be used with Power chips to accelerate particular types of workloads.
A previous version of this story incorrectly stated the name of Suzhou PowerCore.