Advanced Computing in the Age of AI | Tuesday, March 19, 2024

Early Haswell Xeon Buyers Push Intel Datacenter Biz 

For the past several Xeon E5 processor launch cycles, customers in the cloud and supercomputing spaces get their hands on the chips and put them into production systems well ahead of the actual product launch date. This seems to be happening again with the impending "Haswell" Xeon E5 v3 processors from Intel, which had a big uptick in revenues and profits in the second quarter ended in June in its Data Center Group.

The Data Center Group is that part of Intel that makes processors, chipsets, motherboards that are used in servers, switches, and storage arrays. While it is less than half the size of the chip maker's PC Client Group, the Data Center Group, speaking generally over many quarters, is growing faster and brings a larger share of its revenues down to the operating income line in the corporate books. Both are, of course, vital to Intel's health because the chip maker needs PC and other device volumes to help lower the costs of production for server, switch, and storage chips. (And, as IBM's sale of its PC and then its X86 server divisions demonstrate, it is hard for a vendor in the volume server business to do servers without PCs. You need volume leverage to get good pricing from Intel, AMD, and this will be true of ARM chip makers, too, when they get to servers soon.)

On a conference call with Wall Street analysts, Intel's chief financial officer, Stacy Smith, said that revenues for both the PC Client Group and the Data Center Group were better than expected, and that is after Intel had previously revised its guidance upwards. The company booked $13.83 billion in sales in the quarter ended in June, up 8 percent compared to the year-ago period. Operating income across the whole company, which includes software, services, mobile, and other segments, rose 41.4 percent to $3.84 billion. After paying taxes and restructuring costs, net income at Intel during Q2 rose 39.8 percent to $2.8 billion, a cool 20.2 percent of revenues and about as good as anyone can do that is in the manufacturing business. That is the kind of profit margin that a good software company has, in fact.

On the call, Brian Krzanich, Intel's chief executive officer, said that the Data Center Group set an all-time record for sales in the second quarter. Revenues rose a sharp 19.2 percent to $3.51 billion in the quarter. Operating income for the server/switch/storage unit rose by more than twice that rate, up 39.6 percent to $1.82 billion. The other way to say this is that Intel increased Data Center Group revenues by a little over $1 billion, year on year, and just a little under half of that came out as operating income. That is pretty good business if you can get it, and at this point in the history of the computer business, Intel is the only one who can get it. Get it?

You might be wondering how Intel can be setting revenue records with a new server chip in its mainstream line widely expected to be launched at Intel Developer Forum in September. You might expect to be seeing a stall of some sort. Well, as it turns out, early adopters of the Haswell Xeon E5 v3 chips, which Krzanich said would launch later this quarter, are helping to pump up revenues even before they officially come to market, with big cloud builders and HPC customers calling first dibs and getting them, too. The "Ivy Bridge-EX" Xeon E7 processors were launched earlier this year for bigger iron and were ramping through the second quarter, helping boost the numbers, too. Exactly how much of a bump the Haswell E5 and Ivy Bridge E7 chips together gave Intel in the quarter was not revealed, but Intel has said that the overall Data Center Group business can grow in the low double-digits over the long haul and this was clearly better than that. Average selling prices for chips sold in the Data Center Group were up 11 percent, and volumes rose by 9 percent, year on year.

Business was good across the various segments where Intel peddles chippery for datacenters, and Krzanich said that the cloud, HPC, telecommunications, and enterprise segments all had at least a 15 percent revenue increase in Q2 compared to the year ago period, so growth is balanced. Smith added during the question and answer session that the Haswell Xeon E5s were helping the Q2 numbers and that this high growth in sales will probably not repeat itself in the second half of 2014. This suggests that Intel has closed some very big Haswell Xeon E5 deals indeed.

It also suggests that big customers are going to be able to get to the front of the line in a way that was not possible a decade ago and that is, in fact, reminiscent of the early days of the mainframe market, when demand outstripped supply and customers could – and did – trade their position in line for a shiny new IBM mainframe if they were desperate for the capacity. More efficient IT is real money to the cloud builders, where the systems are their business and hundreds of millions in profits are at stake. You can't blame them for wanting to go first or for getting their way, but you can be jealous that the biggest customers are getting preferential treatment that most enterprises will never see. In the modern economy, the aggregation of many small preferential treatments adds up to a sustainable and perhaps unassailable advantage. Right up to the minute where everything changes. .

The important thing for datacenters is that Intel is rolling along with its tick-tock upgrades to Xeon processors, and that even with a six-month delay in the production of "Broadwell" PC chips using 14 nanometer processes, Intel is now shipping the Core-M parts to PC makers and we will see the first devices using them by the holiday season. Broadwell is a process shrink of the Haswell design, what is called a tick in the Intel parlance. Broader availability of a wider Broadwell Core line is coming in the early part of next year. It won't be long before Xeon derivatives of the Broadwell lineup start appearing, including a Broadwell system-on-chip design for microservers expected perhaps this fall. Krzanich said on the call that the Skylake chips, which will have a new microarchitecture – what Intel calls a tock – and will be based on the same 14 nanometer processes as Broadwell, will debut in 2015 and that later this year Intel will start pinning down dates.

When we will see future Xeons based on Skylake remains to be seen, but as Krzanich pointed out, it takes a year and a half on average to roll a tick or a tock from a PC chip up to the fattest server chip, and in the case of Haswell, it is taking more like two years to get from the PC to the Xeon E5 and longer still to get to the Xeon E7. This is probably not fast enough for the hyperscale datacenter operators, and it will be interesting to see if the gap between when they get chips and when they get announced keeps getting wider and wider as the largest chip buyers press for advantage. Instead of overclocking the chip, perhaps they can overclock the roadmap?

EnterpriseAI