A lot has been written about servers based on ARM processors, but there's little data about how many are actually being used in production. That’s still the case, but we now have another large customer that’s at least admitted to trying them out -- and it’s reasons for doing so are not good for Intel.
Morgan Stanley is testing a server with ARM processors from AppliedMicro for use in its data centers, said Bert Shen, vice president of technology business development at the financial services firm. He spoke on a panel organized by AppliedMicro at the ARM TechCon conference in Silicon Valley this week.
The chip maker showed Morgan Stanley a benchmark that compelled it to take a closer look at ARM. "They ran a very relevant columnar database benchmark for us and got a 5x performance improvement per rack compared to an Intel Haswell EP solution,” he said. “We thought this was about the right time to bring in a POC [proof of concept]."
It’s a small test -- Morgan Stanley is running a few ARM cartridges in an HP Moonshot chassis -- but it can help adoption of new technologies if well known businesses go public with their efforts. "We’re excited about what the results might end up being but we’ll have to wait and see," Shen said.
Performance is only one reason Morgan Stanley is looking at ARM. "There's another really important aspect to this, which is our sourcing strategy," Shen said. Like other large companies, it doesn't like being beholden to a single vendor for an important technology.
"For large enterprises, ideally you have two vendors for a product line. You try to avoid three vendors unless it's a commoditized product. But when there's one vendor, you basically run like hell."
AMD is technically a second supplier of x86 server processors, but Intel utterly dominates the market.
It was one of a few signs at TechCon that the market for ARM servers is moving forward. ARM itself, which has tempered expectations in the past, is now more bullish -- it said this week that by 2020 it expects 25 percent of processors going into server sockets to be ARM-based, a huge target to try to reach.
ARM processors are best known for their use in smartphones and tablets, but advocates say they can be more energy-efficient for server tasks that are highly parallelized and can be spread over many cores, such as dishing up Web pages or crunching big data in distributed programs like Hadoop.
Cavium, Broadcom, AMD and Qualcomm are also developing ARM server chips, though AppliedMicro seems to be furthest along. CEO Paramesh Gopi showed the company's latest X-Gene system-on-chip at TechCon.
XGene 3 will have 32 processor cores, running at up to 3GHz, with eight DDR4 memory channels and 42 lanes of PCI Express Gen 3. It will be manufactured using 16nm 3D FinFet transistors.
"This is all going to be available in the second half of 2016," Gopi said, holding a sample chip in the air.
He claimed X-Gene 3 systems will be able to take traditional scale-up workloads and run them on scale-out machines, using a new interconnect that AppliedMicro will announce separately at the SC15 supercomputing conference on Monday.
ARM servers will "no longer going to be confined to workloads that have inherent parallelization," he said.
Hewlett Packard Enterprise and EMC were at the AppliedMicro event, and expressed interest in using ARM processors in future storage systems.
"We fundamentally think choice is valuable," said Michael Robillard, an EMC distinguished engineer. A key selling point for ARM is that the company has many licensees, creating options for buyers.
"We have this stuff in our labs, we've been taking measurements on it and we’ve been impressed with what we’ve seen," he said.
Dave Preston, a distinguished technologist at HPE, said his company tested an X-Gene processor attached to four SSDs and hit a "random read workload" of a little over 200,000 IOPS (input/output operations per second). "It's quite compelling for a low-cost chip," he said.
ARM has one of two architectures trying to challenge Intel's dominance. IBM opened up its Power design last year to let other companies build Power servers. It's part of a broader shift in hardware development that was fueled by the Facebook-led Open Compute Project.
Indeed, the initial target for ARM servers is large-scale cloud providers like Facebook and Google, as well as large enterprises that have the engineering resources to design their own systems and software.
"Maybe the ecosystem isn't as evolved as it should be, but we can evolve it ourselves," said Shen, noting that Morgan Stanley developed its own software container technology years before it was popularized.
The hardware choices around ARM are still limited, however. ARM says there are at least six server models available, but most are development boards from Taiwanese manufacturers. HP is the only major server OEM selling an ARM system. Gigabyte has a server based on Cavium's 48-core ThunderX SOC, but a representative at Cavium's booth said they're not yet in production use.
PayPal and a French hosting provider, OVH, have discussed using ARM servers, but the extent of the deployments is unknown.
Needless to say, Intel isn't standing still; just this week it launched the first SOCs built around its Xeon processors, and its acquisition of Altera will allow it to further customize its server chips for different workloads.
ARM maintains a need for more competition will win out. "What's the value of bringing a competitive market to servers versus a monopoly?" ARM CTO Mike Muller asked reporters Tuesday. "There's an overarching price-competition issue."