Low-power server manufacturer Calxeda on Wednesday outlined a roadmap through 2014 that calls for the company’s fabric technology to tie together chasses and racks into logical clusters.
In 2013, Calxeda plans to introduce “Midway,” which will integrate more robust 32-bit ARM CortexA15 cores. A year later, most likely at the end of 2014, Calxeda’s “Midway” architecture will incorporate the first wave of 64-bit ARM chips, seen as viable competition to the X86 market dominated by Intel.
Calxeda recently announced $55 million in additional funding, evidence that the company is attracting attention from the venture-capital world. Penguin Computing also signed on as a Calxeda customer, developing the UDX1, which features a modular architecture that can be configured with up to 48 Calxeda EnergyCore server nodes (with four cores per node). The system includes an internal 10 Gigabit Ethernet switch fabric for node-to-node connectivity and provides up to 144 TB of hard drive capacity, Penguin said.
The idea, according to John Mao, a product manager at Calxeda, is to dramatically increase the number of addressable cores. “Our current chips support 4,000 cores in a single fabric,” Mao said in an interview Oct. 17. “That’s actually a limitation of our first-generation fabric switch. And our future generations we’re going to be addressing larger sets of single-fabric clusters, into the hundreds of thousands and eventually into the millions, if we can get there, but that’s a bit further out.”
The improvements being made in “Midway” are centered on the core, which adds 50 percent performance per clock, and 3 times the memory bandwidth, albeit at the same frequency as the company’s first-generation part. Additional improvements include hardware virtualization support and larger memory addressability, up to 16 Gbytes of physical memory per chip, or four gigabytes per thread or per core.
Over time, Calxeda plans to rethink how it can arrange its cores. “We’ve been trying to see if there’s a native market demand for redefining what we traditionally think of as the chassis,” Maio said. “In these very large scale data centers, they don’t buy individual chasses, individual servers, they buy racks at one time. The minimum purchase order is ten racks, or a dozen racks.”
He added: “If you’re buying at that scale, what if we rethought what a traditional chassis would look like. What if you were able to build—and this is all theoretical—what if you built an entire rack… as a single chassis?”
Inside of that highly theoretical rack, the company could leverage its technology to streamline its clients’ overhead. Let’s say there are 4U chasses. How do you effectively and logically make that appear like one entire rack-based cluster without being spending on traditional infrastructure like networking overhead and so forth? So that’s kind of our approach.”
Image: Supri Suharjoto/Shutterstock.com