Dell Sees Server Uptick, Awaits Demand For ARM 

Dell is no longer a public company, so we can't look at its quarterly financial results to get a sense of how its various business units are doing. But, Dell is a major IT supplier and it wants to let customers and partners alike know how things are going since it became public last November, and to that end Forrest Norrod, vice president and general manager of Dell Server Solutions, had a chat with EnterpriseTech about the state of the server business.

"The market is pretty interesting," says Norrod. "It really feels like we hit an inflection point at the beginning of the year where enterprise started to invest again. When we look at it, we still see growth but it is moderating in terms of the percent for the hyperscale and cloud guys. But the thing that still feels very different is the demand on the general enterprise space is markedly different from the last two years.”

This is consistent with what Intel saw in its first and second quarter in 2014 as well. In the quarter ended in June, Intel's Data Center Group, which makes chips, chipsets, and motherboards for servers, switches, and storage, had a very big 19 percent revenue bump, driven in part by steady "Ivy Bridge-EP" and early "Haswell-EP" Xeon E5 server sales as well as the ramping of high-end "Ivy Bridge-EX" Xeon E7 processors for machines with large core counts and memory footprints. Intel's numbers were no doubt helped by hyperscale datacenter operator and supercomputing lab customers who got early access to the Haswell parts.

"I don't think it is all Haswell," says Norrod, referring to Intel's growth in Q2. "There have been some very large Ivy Bridge deals in the past three months as well."

Dell carves up the server market and attacks it in a slightly different fashion from its peers and Norrod says that each of the segments it builds machines for are seeing growth.

"We typically think in big animal terms," Norrod explains. "The true hyperscale market is a very small set of customers, maybe the top seven to ten players. The scale-out customers sit below these, and include Web tech, HPC, and the large financial institutions for their quant farms. The core enterprise comes next and includes converged, high-value workloads and volume workloads, and finally there is the SMB/value segment. All four of these segments are growing right now. The strongest unit growth is probably on that scale-out space below hyperscale and we are still seeing great opportunities for Web tech and technical computing. I think that HPC is becoming less and less a thing off in the corner and more of a critical component of almost everybody’s business. And the interesting thing from our perspective is not necessarily the exascale ambition and hundreds of millions of dollars in government projects. We are much more interested in the commercial, mid-scale, and educational technical computing areas and we think these are fast growing segments. Core enterprise has returned to growth.”

Generally speaking, the scale-out part of the Dell server business has the most growth, followed by converged infrastructure in the core enterprise, which is followed by hyperscale machines, and then the balance is the core enterprise segment.

The Data Center Solutions unit of Dell, which was once building custom servers for Facebook before the latter company founded the Open Compute Project to design its own servers, storage, switches, and datacenters, used to be around a $1 billion-a-year run rate across a handful of customers and for the past several years, Dell has been making quasi-custom machines and also pushing custom iron down into smaller but equally innovative hyperscale and cloud companies. The company no longer talks about this business separately, but more as one end of a spectrum of services it offers to provide anything from general purpose to custom machinery. That hyperscale business has cooled a bit.

"In terms of unit shipments, the hyperscale business is still growing in the low double digits," says Norrod. "It is no longer at the triple digit or high double digit growth that we once saw. I think that it is just the limit of large numbers if nothing else. It is feast or famine on the quarterly numbers and sometimes on the annual numbers, but overall, but if you iterate over a couple of quarters, the hyperscale market is still showing low double digit growth."

As far as market share goes, it is hard to gauge what is happening among the true hyperscale powerhouses, who by and large design their own gear and farm out manufacturing to original design manufacturers (ODMs). Norrod says that if you take data for density-optimized servers and then ODM/direct sales and use them together as a proxy for the scale-out portion of the customer base, which the analysts at IDC track, then these two areas constitute about 20 percent of the unit shipments, give or take, and a little bit lower share of server revenues. Dell has held onto the number one position in the density optimized part of the market, based on IDC data, except for during the second half of 2013, when Hewlett-Packard was particularly aggressive and won a big deal at a large account where HP had been losing for years. “I think that they had their time in the sun and it has turned back to what it had been for years, if you catch my drift," Norrod says.

The obvious question is: Will this pattern of sales in the server business at Dell hold? Will something – like the advent of ARM servers in the datacenter, or a resurgence in Power chips – somehow change the pattern?

“I am always skeptical about trying to predict two or three years out into the future," Norrod says. "I think that anybody that tells you what is going to happen in three years is either lying or delusional or both. But if I think about big macro trends, then there is still a lot to be done around scale-out for both HPC and analytics, and from the point of view of infrastructure they sort of lump together in the same space. On the core enterprise side, I think more and more interest in converged and even hyperconverged systems like those from Nutanix and its competitors. We see a lot of growth there. People have virtualized and they have got the utilization, but they still haven't been able to get the agility. Hyperconverged gives them a complementary answer to the public cloud for on-premise equipment."

The difficulty of predicting the future is apparent when it comes to ARM servers. Dell has been way out in front when it comes to experimenting with ARM chips in server designs, including a set of hyperscale machines code-named "Copper" that were announced back in May 2012 using Marvell's 40bit ARMv7-A-compliant processors and the "Zinc" machines that were based on the now-defunct 32-bit Calxeda EnergyCore 1000 processors. Dell has been mum about its plans for AMD's 64-bit "Seattle" Opteron A1100 and AppliedMicro's 64-bit X-Gene1 processors, which are sampling now and will ramp throughout the year. Cavium and Broadcom are working on their own server variants of the ARMv8 design, but Nvidia and Samsung Electronics are apparently bowing out of the server space with their ARM efforts.

With Calxeda having gone out of business late last year and the launch of 64-bit ARM chips and software stacks from them taking longer than expected, the hype has died down for ARM servers a bit, and the expectation, says Norrod, is for the market to start forming in 2015 and develop further in 2016 as 64-bit chips ramp.

“We see the benefits and gains that ARM can offer customers when it comes to performance, power efficiency, and cost and we support fostering that ecosystem," he explains. "Ultimately customer needs dictate what we offer and right now our mainstream server customers want versatility and robust software tools that X86 developers have long enjoyed. We anticipate the space to become increasingly competitive over the next few years and we’ll continue to stay engaged to best enable our customers and their evaluation of ARM systems."

As for Power machinery, Norrod says that IBM is "clearly trying to pivot" those HPC and enterprise customers who had been buying its X86 clusters or BlueGene massively parallel machines to clusters of servers based on its Power8 processors. Eventually, IBM will offer a variety of coprocessor accelerators for these machines, aimed at speeding up applications of all kinds both in enterprise datacenters and supercomputing centers.