Advanced Computing in the Age of AI | Friday, May 20, 2022

Intel Datacenter Biz Profits From Hyperscale, HPC Build Outs 

In most cases, saying the datacenter is hot is a bad thing, but not so when talking about the financial results of chip maker Intel. The company just reported its financial results for the third quarter, kicking off the tech earnings roll call, and the news was better than expected and sales of processors, chipsets, and motherboards in the Data Center Group are sizzling.

In the quarter ended in September, Intel's Data Center Group reported $3.7 billion in revenues, up 16.4 percent compared to the year-ago period. Because the Xeon and Atom server chips are largely based on existing 32 nanometer and 22 nanometer processes that are mature and customers are trending up toward more expensive – and more profitable – variants of its processors, the operating income for the enterprise portion of Intel rose by 26 percent to $1.92 billion. It didn't hurt that Intel was also ramping up its "Haswell" Xeon E5-2600 processors during the quarter, either. These Haswell Xeon E5 chips were formally launched on September 8, but Intel had begun shipments to early customers in the cloud and high performance computing markets several months earlier.

On a conference call with Wall Street analysts going over the numbers, Intel CEO Brian Krzanich said that the Haswell Xeon E5 processors represented about 10 percent of the Xeon volumes for dual-socket servers already. The two-socket server represents the lion's share the platforms using Intel's Xeon processors.

Intel does not call out its Xeon and Atom processor sales separately or by server type in its financial analysis, although it does offer some segmentation for the PC Client Group for notebook and desktop platforms. But Intel did say that unit volumes for processors in the Data Center Group were up 6 percent year on year in the third quarter, and that average selling prices were up 10 percent. This is all the more remarkable considering that Intel is pushing low-end Xeon and Atom processors into more storage and network devices and, to a certain extent, into microservers and other relatively low-powered (in terms of both performance and electricity usage) systems. Clearly, virtualization, clouds, and hyperscale datacenter operators are buying beefier processors, driving up the overall ASP.

While Intel said it had broken through the 100 million CPU barrier for the first time during the third quarter, the company did not break out datacenter and client CPUs separately from these numbers. But client devices represent the majority of those shipments for sure. Somewhere between 2.2 million and 2.5 million servers ship in any given quarter these days, and these are for the most part two-socket machines. Call it 5 million CPU units in total, with the dominant shipments being for X86 machines. There is some wiggle in these numbers because Intel is shipping into ODMs and OEMs and directly into some customers, and IDC and Gartner count the machines that ship to customers. Intel is talking about the water coming out of the faucet and the server box counters are talking about the water exiting the bathtub; depending on the flow rate (which is largely gauged by the worldwide economy's gross domestic product among enterprise customers), there can be more or less water – or rather, chips – in the tub at any given time.

Giving us a little peek into the Data Center Group business, Krzanich said that Intel is now shipping 35 custom processors to customers, which sometimes have instructions only accessible by them, sometimes run at higher clock speeds or temperatures than off-the-shelf parts, or have other different packaging that Intel has not specified. The volume shipments of these custom processors – we presume them to be largely custom Xeon parts, and ones with higher clock speeds and thermal ratings – are growing three times as fast as the off-the-shelf parts, Krzanich said. One concern with such mass customization is that it may eat into profits because there is a cost to such custom engineering and manufacturing. But there is also a cost to not giving customers what they want, too, and it is clear that the hyperscale datacenter operators and to a lesser extent HPC centers and large financial institutions are able to get what they want in terms of customization to goose their specific applications without hurting Intel's middle and bottom lines. As Intel gets better at this mass customization, it is better able to counter a possible threat from the myriad ARM server chip makers that want to take a bite out of its Data Center Group business.

Breaking down Data Center Group sales in the four dominant segments that Intel tracks, revenues for products that ended up being sold to build enterprise server products rose by 11 percent, and sales of components for networking products rose by 16 percent. Sales of components for HPC systems rose by 22 percent, and those for cloud and service providers – Intel lumps what we call hyperscale datacenter operators in here – were up by a staggering 34 percent. This should portend an uptick in server revenues, but perhaps not. Remember, the trend is for increasingly minimalist server designs specifically so customers can spend more money on processors and memory. The server ASP does not have to rise just because the CPU ASP does, and it is better for Intel, in fact, if it does not.

Earlier this year, Intel was cautiously optimistic that enterprise server spending would rise, and Krzanich reiterated the company's expectation that Data Center Group would grow revenues by about 15 percent this year. "We are expecting Q4 to kind of progress from Q3," Krzanich said on the call. Intel is hosting its investor day on November 20, but is thus far not providing detailed projections for 2015 except that it believes that Data Center Group can grow at about 15 percent per year going forward. In 2013, Data Center Group posted sales of $11.2 billion, up 6.9 percent, but operating income only rose by 2.9 percent to $5.16 billion. It looks like for 2014, Data Center Group will kiss $13 billion in sales and if present conditions persist, then Intel could break $20 billion in revenues for the group – a goal it has had for many years – by 2017.

Without a credible threat from alternative CPU architectures – ARM has yet to prove itself, and Power is better positioned but still miniscule compared to X86 – and assuming the global economy remains stable and Intel continues to have a manufacturing lead, it is hard to imagine a world where the X86 chip – and specifically the Xeon – does not continue to dominate the datacenter.

Across the whole business for the quarter, Intel posted sales of $14.6 billion, up 7.9 percent, and net income of $3.32 billion, up 16 percent.

Add a Comment

EnterpriseAI