Applied Micro Circuits, a chip firm that designs silicon parts for the computing and networking world, has spent the last three years making a big bet on the cloud computing market and the ARM architecture. The results began shipping last week, and the product essentially takes networking and computing and crams it all onto one system on a chip.
Dubbed the X-Gene server on a chip, the product has been touted by Applied as the first 64-bit-capable ARM-based server in existence, the ideal part for webscale users (check out the pic of Facebook’s Frank Frankovsky holding one up) and also the future of Applied Micro. It’s the first chip to contain a software-defined network (SDN) controller on the die that will offer network services such as load balancing and ensuring service-level agreements on the chip. It’s like shoving the networking and computing vision of the Cisco Unified Computing System on a chip.
This is a big deal. Although the first generation won’t have enough bandwidth to eliminate the need for a switch at the top of a rack, the following generation will.
Paramesh Gopi, president and CEO of Applied Micro, said that these new chips have now made it past the prototype stage (the board in the picture uses an FPGA instead of a production silicon) AND are now in the hands of several customers, including Dell and Red Hat. Gopi expects physical servers containing the X-Gene to hit the market by the end of this year.
Gopi’s big bet
The chip is manufactured at 40 nanometers and contains eight 2.4 GHz ARM cores that Applied has designed, four smaller ARM Cortex A5 cores running the SDN controller software (the pink bit on the block diagram below), four 10-gigabit ethernet ports, and various ports that can support more Ethernet, SSDs, accelerator cards such as those from Fusion-io or SATA drives. In short, this a chip that combines networking and computing in one package.
When about asked about the power consumption of the chip, Gopi said it will run at 50 percent of the total cost of ownership of a comparable x86 product, but wouldn’t discuss actual power consumption.
“We’ll be able to run your LAMP stack and SQL jobs on Xeon-class ARM cores, and the routing protocols and such will be running on the Atom-class ARMs,” Gopi said. “It’s the fundamentals of a rack on a single chip.”
Building this chip has taken four years. It required Gopi to visit ARM at its U.K. headquarters to convince them to give him an architecture license to build a chip for servers. In an interview with me at the Open Compute Summit in January, Gopi explained that he saw the flexibility and the architecture that ARM offered could become an asset for webscale computing, so he embarked on turning Applied Micro, a public company with a few hundred million in revenue, into a startup.
Like others, such as Barry Evans of Calxeda or Andrew Feldman of Sea Micro, he saw that power issues were raising the cost of operating data centers — and cutting into the bottom line at web businesses — and he thought he had a solution. His solution was to get an architectural license from ARM, so he could make a 64-bit-capable chip ahead of ARM’s plans to introduce that powerful a core. ARM introduced that core last year, and vendors of ARM-based server chips such as AMD and Calxeda expect to have 64-bit-capable chips next year. But Applied is shipping those machines today.
“We’ll end this wimpy core vs. brawny core debate once and for all,” Gopi said.
The new hardware mindsetGopi has taken advantage of several different trends that are finally coming to fruition. The first trend is the use of the ARM core — ubiquitous in cell phones and tablets — for the enterprise and cloud computing market. But he’s also taking advantage of a more subtle shift happening in the chip world as it pertains to the data center — namely the opening up of the ecosystem.
The mobile industry has relied upon the common ARM architecture to build a wide variety of chips that give each vendor a slightly different set of features. Both Nvidia and Qualcomm start with ARM cores (hell, even Apple has an ARM architectural license) to build their application processors. This lowers the cost of designing chips, because engineers can start from a higher level when solving problems.
And the modularity of the ARM cores combined with an architecture license also means firms can customize their designs for a certain market without spending a huge amount of time or dollars. Gopi will actually address some of this at our Structure event June 19 and 20, in a presentation on designing hardware at the speed of software.
For Applied, this dynamic plays out in the existence of a new type of chip for the data center, but also in the fact that in nine or 12 months Applied plans to test the second-generation X-Gene chip, one that will support 100-Gigabit Ethernet and will obviate the need for a top-of-rack switch. Ironically, this architecture probably won’t be a welcome development for Applied’s existing networking clients like Cisco and Juniper.
But it’s clearly the direction that large webscale customers want to go. And the second-generation architecture is also important for the first-generation X-Gene products, because without it, Applied may not have a chance at getting technically savvy and forward-looking potential customers that need not just a single interesting product, but a real understanding of the roadmap before they commit to a new architecture.
So even as Applied ships these first products to customers for use in devices that hit the market at the end of this year, it’s already developing its production of the next generation 28-nanometer versions of the heavy-duty ARM cores and 100-Gigabit-capable networking while prepping for later versions that may include photonics and other elements that data center customers are already discussing as tomorrow’s technology.
It took a bold vision — and that trip to ARM — for Gopi to get Applied Micro to the table as these discussions about the next generation data center are playing out. But with this design, it has earned a seat. Now all it has to do is earn the business.