Confirmation of the partnership between Hewlett Packard Co. and Texas startup Calxeda Inc. to develop ARM-based servers has added more fuel to discussions about microservers, and which process technology is best poised to serve the market.
On Tuesday (Nov 1), HP said it would be working with Austin-based Calxeda on a development system for ARM based server offerings, allowing customers and partners to initially test systems using Calxeda and later other ARM and x86 chips
Last week, ARM Holdings CTO Mike Muller announced his firm’s upcoming v8 architecture and first 64-bit instruction set to effectively push ARM architecture into new segments of both the consumer and enterprise markets, including the server space.
ARM’s v8 will support both 32-bit and 64-bit applications, said Muller, addressing a major issue in the firm’s server space aspirations. Intel and its supporters have long argued that without 64-bit, ARM would simply lack the memory needed to support legacy software.
“ARM’s announcement is proof that 64-bit is essential,” an Intel spokesman told EE Times, adding, “without it, ARM’s hopes to enter the server space would have been challenging and difficult.”
ARM has long been toying with the idea of having dense, ultra-low power servers based around its architecture, targeted at the exponentially growing mega-datacenter market serving the cloud.
Indeed, there is an argument that clouds hosted by companies such as Amazon Inc., Google Inc. and even Apple Inc. don’t need as much compute power as they do lower-power consumption and the ability to fit more compute engines into a smaller space. After all, the applications that run on the aforementioned clouds are fairly simple and can largely be parallelized.
One could argue, as HP and Calxeda are now doing, that even the current lack of 64-bit ARM architecture should not prevent firms from picking these new non-x86 microservers for certain applications. HP has even gone on record to say it does see a role for 32-bit ARM servers in the market.
The newly announced HP “Redstone” development system, which swapped out x86 in favor of 18 10-inch by 3-inch Calxeda EnergyCard servers, sporting four SOC quad-core server nodes, purportedly needs just 5W per server and 20W per card maximum power draw. In addition, each server tray of 72 quad-core ARM servers takes up just one rack unit equivalent of space.
Those two factors alone, and the idea of giving throughput per Watt such a boost, make the HP/Calxeda project interesting enough for evaluation as an alternative when it comes to HPC and hyperscale web computing, 64-bit compatible or not.
Intel, however, is not likely to feel immediately concerned. After all, it too is moving forward rapidly and even coined the term “microserver” with its announcement two years ago that it would start making server products using low power versions of its 64-bit Xeon and Atom chips.
The chip giant has not only published a microserver specification, but runs its own development lab for microservers and has customers like Dell, NEC, Hitachi, and Super Micro already selling its low-power server products to customers. “We are currently the only company in the microserver market,” Intel’s spokesman said.
More efficiency from 14-nm
By 2014, when the first 64-bit ARM cores start surfacing in servers, Intel’s Atom roadmap will already be at the 14-nm mark, bound to offer significant power advantages over current Atom models.
“The manufacturing process will certainly come more into play,” said Intel’s spokesman, adding that 14-nm would deliver “a huge kick across all vectors of power efficiency.”
Then, of course, there is also the question of software compatibility to be taken into account.
“The biggest proof-point for ARM and its partners will come down to the software,” said In-Stat analyst Jim McGregor, explaining that what mattered most to server customers was whether the product would work without emulating and without “causing pain” to those having to make the switch.
“People are still using IBM mainframes based on software decisions,” he said. How ARM and its partners managed to deal with the software conversion process would be critical to its success in microservers, said McGregor. “Software migration can be the biggest cost and hurdle for customers, so ARM has to get that right,” he emphasized.
Intel also has years of 64-bit experience under its belt.
“Time works to our advantage as we accelerate the Atom roadmap,” Intel’s spokesman said, adding that the software ecosystem argument would remain valid until ARM could prove itself able to handle server software and legacy applications.
Another potential challenge for ARM’s server aspirations could be a lack of concrete strategy, said McGregor. McGregor posited that ARM may have been pushed into announcing 64-bits before it had a solid plan in place, noting the announcement had come somewhat out of the blue.
“I got the feeling several key customers drove this decision,” he said, adding that both Apple and Nvidia may have fallen into that category.
Apple, McGregor pointed out, actually had an acute need for servers and had dabbled in the market in the past. “Apple is not just a consumer electronics company, it’s a data center company,” he said, pointing out that the Cupertino-based firm had made multiple investments--as well as an acquisition--in the process technology space.
McGregor also believes ARM may have made the announcement without knowing which specific core 64-bit was intended for. “I almost think ARM came out with this before it had any actual cores in mind,” he said, adding that his questions to the firm pressing for more information had thus far gone unanswered.
The Apple hypothesis does tie in with other analyst speculation about the market for microservers. With their lower power and space saving merits, such machines would be perfect to host services like iTunes or the iCloud.
Google, too, would find a use for such servers, notes Roger Kay of Endpoint Technologies in his analysis.
“Large but sparsely used databases are just perfect for this type of machine. If Google buys a big pile of these systems, which stack and aggregate rather nicely, it can save a lot of money on both purchase and operating costs, lowering its power bill while packing more servers into the same floor space,” Kay wrote recently.
According to some estimates, densely-packed servers that address large parallelizable datasets and simple queries could make up to 10-15 percent of the entire server market, representing about $6-9 billion of the total addressable market, a segment certainly worth pursuing.
“The microserver industry is still very young and a lot more work still needs to go into its development,” Intel’s spokesman said, adding that ARM’s announcement on 64-bit had always been expected.
“We even thought the announcement might come earlier, because 2014 seems a long way away,” he said. Then again, with HP’s Calxeda initiative, ARM might not need to wait three more years to start making its presence felt in the microserver space.