Applied Micro lifted the curtain on its 64-bit ARM-based server platform at the Hot Chips conference this week and ignited a debate about whether ARM is the right architecture for the data center.
Pat Gelsinger, a former Intel CTO who is now president and COO of EMC, said he doesn't think ARM chips are right for servers. "I don't think the math makes sense," he said.
Applied disagrees, contending that for certain workloads, including big data and Web-scale cloud-computing applications, its server chips will lead to dramatically lower ownership costs, largely through reduced power consumption.
Applied said it's on track to ship samples of its system-on-chip, called X-Gene, in the second half of the year. It will eventually be offered in servers with 128, 256 and 512 processor cores, President and CEO Paramesh Gopi said in a presentation Wednesday.
Several vendors are working on server products based on 32-bit ARM designs, but Applied may be the first to release a server product based on the new ARMv8, whose 64-bit architecture is more suited to server workloads.
According to Applied, integration is key. Its SOC will be offered with eight or 16 CPU cores, packaged on the same silicon die as the I/O, networking and interconnect fabric. That's built into a small server module with up to 256GB of DDR memory.
"This has never happened in the history of our industry," Gopi said. "For once we're not talking about CPUs, we're talking about a complete server-on-a-chip platform, and that is going to change the fundamental equation" for the cost of server ownership.
Applied won't say yet how much power its platform consumes or discuss the speed of its CPUs, saying product announcements will follow. And while Gopi held up a server module on stage and declared dramatically, "This is real," another Applied engineer admitted later that the SOC was a mock up, since the company is still awaiting its first samples.
Still, Applied said it's on track to deliver the first parts by the end of the year, and it says they'll run software written for 32-bit ARM devices without compromising performance. It's platform will also support virtualisation and enterprise-class RAS (reliability, availability and serviceability), it said.
Earlier in the day, Gelsinger said that while the shift to smartphones and tablets means ARM will dominate client computing -- "I don't see that changing any time soon, with all deference to my x86 roots" -- he believes the data center will remain "an all x86 world."
Asked what he thinks of analyst predictions that ARM will account for 25 percent of server shipments a few years from now, Gelsinger said he doesn't buy it. "I don't believe those numbers," he said flatly.
Processors based on the x86 architecture are getting to as low as 2 or 3 watts per core, he said. Reducing the power of the CPU any further doesn't make sense when taking into account the power needed to run the other components in a server platform, such as the I/O and memory, he argued.
"I don't think the math makes sense," he said.
Add to that the overhead of trying to virtualise and manage a data center with a mix of ARM and x86 servers, and ARM makes even less sense. "The cost of heterogeneity in a cloud computing model is very high," he said.
Applied disagreed. Gelsinger's argument holds true for traditional x86 servers that have their own individual chassis, power and cooling supply, but the industry is heading toward a shared infrastructure, and even shared components such as I/O and memory, said Gaurav Singh, Applied's vice president for engineering and product development
"In the new systems, the math makes a lot more sense," he said.
Analyst Nathan Brookwood of Insight64 sided with the ARM camp. "Why I believe ARM makes sense is because it allows you to innovate at the CPU-level in a way that you can't with x86, because Intel and AMD are the only guys who have access to those cores," he said.
Applied, by contrast, was able to build a highly integrated SOC by innovating around ARM's CPU core.
SeaMicro, the server startup bought this year by AMD, has been selling even denser servers than Applied is proposing, based on Intel's Atom and Xeon processors. But SeaMicro had to use a two-chip solution, with its networking fabric on a separate piece of silicon, because it couldn't modify Intel's chips, he noted.
Now that it owns SeaMicro, AMD has hinted that it will integrate the company's technology onto its own Opteron server chips. "But that's a couple of years away," Brookwood said.
Gelsinger isn't an unbiased observer, of course. In two days he's scheduled to take over as CEO of VMware, which is mostly owned by EMC, and VMware has built its business on the x86 architecture. It's software, at least for now, does not virtualise ARM servers.
But while Gelsinger has serious doubts about ARM, he's apparently not religiously opposed to it. "If they emerge, if they're really good, of course we'll make our software support them as well, but I'm very sceptical," he said.