Revenues in the quarter ended in March were up 38.8 per cent, to $199.4m, but that was less revenue than expected and rising costs on all fronts plus another $19m in restructuring costs pushed SGI to a loss of $1.2m (still better than the $1.7m loss a year ago but clearly not what SGI had been planning for).

Across all revenue, the public sector drove 61 per cent of sales in fiscal Q3, compared to 19 per cent for hyperscale cloud operators (Amazon is SGI’s biggest customer in this area) and 8 per cent for manufacturers. Cloud revenues were up 15 per cent in the quarter thanks to a big order, Titinger said on a call with Wall Street analysts going over the numbers.

Titinger and Jim Wheat, SGI’s outgoing CFO, have identified nine deals worth a combined $87m that will book in calendar 2012 that have low margins that it has to ride out; most of the deals are for its ICE X X86-based clusters, Titinger said, and the deals have single-digit margins, which he was obviously not happy with.

“In the past, we focused on the top line and growing key customers, which has led to price pressure in some instances,” explained Titinger. “The side effect has been an adverse effect on our margins, which will be in effect until the end of the calendar year. Therefore we are implementing a more rigorous deal review process.”

The other thing that SGI will be looking at is what it can do to get paid a little quicker, Titinger said, because on some big deals it can take three or four quarters to get machinery qualified and accepted by customers.

SGI is in the middle of restructuring its European operations as well, a process that will continue until the end of the calendar year but which should make SGI’s EMEA operations profitable by the end of fiscal 2013 as it eliminates around $7.5m in costs from the region.

Wheat said on the call that SGI, under his successor CFO Bob Nikl, (hired at the end of April) would be giving quarterly guidance from here on out instead of annual guidance as SGI has been doing for the past couple of years. To that end, SGI warned Wall Street that while it will begin shipping its next-generation UV2 massively parallel supercomputers, presumably based on Intel’s forthcoming Xeon E5-4600 processors (which are expected soon) and a rev on the NUMAlink 5 interconnect created by SGI, in the current quarter, sales of the first-generation UV machines took a hit because everyone is waiting to see what the new machines pack in terms of punch.

Moreover, Intel’s Xeon E5-2600 processors, which were announced in March, were a few months later than expected (something on the order of four to six, depending on who you ask) and that had an adverse impact on the ICE and Rackable system sales in fiscal Q3. All of this affected fiscal 2012 revenues and profits.

And thus SGI say that it anticipates sales of between $177m and $197m in the fourth quarter with a loss per share of 56 cents to 71 cents. For the full year, SGI is guiding to between $750, to $770m in revenues, down $20m at the low end and $30m at the high end from its previous guidance for fiscal 2012 and is happening in part because a large deal expected in Q4 fiscal 2012 has slipped into Q1 fiscal 2013.

The company now expects to lose somewhere between 75 cents and 90 cents a share for the full year, which is a lot deeper than the 15 cents to 30 cents loss Wall Street was expecting based on the previous guidance.

SGI has over 600 patents and its own supercomputing interconnect, the former of which you can bet Titinger is looking to peddle on the open market and the latter of which it may feel compelled to do, too, as Cray has done.

The plan is for SGI to be profitable on a non-GAAP basis in fiscal 2013, Titinger was squeezed a little to say on the Wall Street call, but he did not provide revenue or profit guidance beyond that. He expects to have his company review done soon and a battle plan ready for the call in August going over the Q4 fiscal 2012 numbers.

In a separate but no doubt related announcement, the US Department of Defense is shelling out $27.8m to upgrade the supercomputing facilities of the Air Force Research Lab as part of its High Performance Computing Modernization Program. The deal involves the Air Force installing a 32-rack ICE X cluster with 2,304 half-width, double-stuffed “Gemini” IP-115 system boards.

These blade servers are designed to mount one on top of the other – it’s more snuggling back to belly than missionary, so you get the right idea – with a total of 9,216 Xeon E5-2600 processors from Intel running at 2.6GHz for a total of 73,728 cores. This box will pack a 1.5 petaflops peak theoretical performance punch. The Air Force is also getting InfiniteStorage arrays crammed with 6.72PB of capacity for feed this beast.

The Air Force Research Lab currently has a fairly new Cray XE6 super with 2,732 nodes and 43,712 Opteron cores on its current “Raptor” cluster, which weighs in at 410 teraflops. The lab also has an Altix 4700 Itanium-NUMAlink 4 cluster, called “Hawk,” rated at 59 teraflops. This machine is a bit long in the tooth and it is interesting that the Air Force has not upgraded it to UV1; maybe it will get a UV2 machine, maybe not.

The Air Force has another cluster based on Appro International’s Xtreme-X blade server design and using Opteron processors from Advanced Micro Devices as well that is rated at 27 teraflops, which is called the Utility Server in a very un-Air Force way. ®

This article originally appeared in The Register. It appears here in its entirety as part of a cross-publishing agreement.

The move is something of a surprise to us outside of Cray, Intel, and the government supercomputer labs that pay the bills at the supercomputer maker. But Intel has been very serious about networking in the past year and in exascale computing in particular and is seeking every advantage it can get.

Cray added that it would have “opportunities to leverage important differentiating features of certain future Intel products” and would retain certain rights to use the transferred assets and intellectual property in Cray products. (This is a pretty neat trick, considering that Uncle Sam pretty much paid Cray to create the last several generations of machines and then let it commercialize them. This is exactly how the phone companies ended up owning the Internet.) Cray expects the deal to close quickly, and hopefully before the end of the current quarter, and that following the close, 74 employees would move over to Intel.

This agreement is evidence of the leadership position we’ve established in high performance computing, and is an exciting win for our customers, our company and our shareholders,” said Peter Ungaro, president and CEO at Cray, in a statement. “By broadening our relationship with Intel, we are positioned to further penetrate the HPC market and expand on our industry-leading technologies in support of our adaptive supercomputing vision.”

This is certainly a feather in the cap of Diane Bryant, the new general manager of Intel’s Data Center and Connected Systems Group.

“Delivering continued advancement in high-performance computing, including breaking the exascale barrier, requires tremendous innovation in interconnect technology,” said Bryant in Intel’s statement on the deal.

“The acquisition of Cray’s industry-leading interconnect technology and expertise provides exceptional strategic assets that further enhance Intel’s HPC portfolio. We’re excited about the value this will allow us to bring to our customers.”

Intel was talking up how the “Aries” interconnect for the future Cray “Cascade” supers would integrate with the Xeon chips, but to be honest, what it really does is integrate with the PCI-Express 3.0 bus, so technically speaking, the Aries interconnect is less linked to the Xeon than the Gemini interconnect was with the Opterons from chip rival Advanced Micro Devices.

The “Gemini” XE interconnect hooked into the HyperTransport point-to-point ports on the Opterons and has locked Cray into AMD chips – and its supply problems and occasional bugs – for the past decade, causing it grief. Cray broke free of HyperTransport, and did not get sucked into the similar QuickPath Interconnect on Xeons, so it could create an interconnect that would work with any processor, coprocessor, or accelerator.

It is not hard to understand why Intel wants the Cray interconnect. What is hard to understand is why Cray is willing to let go of it at all.

Perhaps Cray has a whole different idea about how to get to exascale, or perhaps Cray is tired of wrangling big development contracts out of the US Defense Advanced Research Projects Agency, which it has done for the current Gemini interconnect for the Opteron-based XE6 and XK6 supers and the future Aries interconnect used in the Xeon-based Cascade machines. But those interconnects are the heart of the Cray systems and if Cray is not going to be an interconnect maker and peddler, then it is reasonable to ask what Cray will be. Perhaps Super-Duper Super Micro Soft?

It could be that the DARPA gravy train has gone off the rails and Cray wants to focus on system integration and other aspects of the supercomputer, like the operating environment and compilers, and leave the future of high-speed interconnects to Intel. After buying Ethernet ASIC maker Fulcrum Microsystems last July for its 10GE and 40GE chips for switches and routers, and then snapping up the InfiniBand biz from QLogic for $125m back in January, it is pretty clear that Intel was going to take a direct interest in networking as a way to get networking better integrated with its Xeon processors and “Knights” family of parallel x86 coprocessors.

So it is pretty obvious why Intel would want the Cray assets, and it is really only paying pocket change for them compared to its own massive R&D budget, which was $8.37bn last year. But for Cray to give up interconnects is just a little tough to swallow, unless Intel has made it clear that it wants to control this and will do it with or without Cray’s help.

Another interesting possibility is that Intel is really interested in building up a patent portfolio in networking that it can use to defend against its new networking rivals and with which it can attack them same. For instance, imagine what might happen if Intel thinks that the SeaMicro 3D torus/mesh interconnect that AMD was so eager to get its hands on infringes on Cray patents? Even if the deal has nothing to do with patents, Intel can now say it has a much more scalable interconnect than AMD.

Cray will be hosting a call at 8 am Eastern on Wednesday (April 25) to go over the deal in more detail. Stay tuned. ®

This article originally appeared in The Register. It appears here in its entirety as part of a cross-publishing agreement.

 

A bang-for-buck comparison

The table above is sorted in performance order. The yellowish rows represent the Apple iPad and some systems I have hanging off my office/home network. The astounding part of the table is the final column – the cost per MFLOP/s. What jumps out is that my wife’s generic business desktop computer kicks ass from a price/performance perspective. At less than 5 cents per MFLOP/s, it’s well ahead of everything else on the list.

The Lenovo W510 was billed as a “Mobile Workstation” when I bought it, and it’s the fastest laptop I’ve ever used. It’s also the heaviest and the most power-hungry, but with an Intel Extreme edition i7 mobile processor and a discreet NVIDIA Quadro graphics processor, it can handle the video processing I throw at it when I’m on the road.

A Bugatti of my own

The Hydra-1 system is something I’ve been working on for months, and it’s finally ready to come into service. I’m getting old enough to realize that I’ll never own a Bugatti Veyron ($2.4m) or even the much less expensive Ferrari Enzo ($670,000), but I can have the fastest computer in my state (until someone proves me wrong – plus it’s a small state).

I built Hydra-1 over the past several months, and it’s quite the screamer. A very helpful HPC vendor helped hook me up with two Intel Xeon 5690s which, together with 24GB of RAM, turned in a Linpack of 122.68 GFLOP/s at stock clock frequencies. Hydra-1 has some serious liquid cooling built in, so I have plenty of room for over-clocking, which might take the Linpack to 140 or better. I think I’m way under the Linpack theoretical max for the system, but I’m running the benchmark under Windows 7 without any tuning, so I’ll take what I can get.

The liquid cooling almost drove me out of my mind, but will definitely pay off on performance and comfort fronts. I’ve written some blogs about the whole process that I’ll submit to The Reg soon.

The iPad does okay, but not nearly as well as the typical systems in the chart. And the iPad numbers don’t include a 3G wireless plan or a cover – and the cheapest poly cover at the Apple store costs $39, or about 6 per cent of the total system price. And if you go with leather? At $69, that’ll drive your price per MFLOP up by 10 per cent.

The biggest surprise on our chart above? Look at Roadrunner and the K computer. Their cost per MFLOP/s is a mere 10 cents and 13 cents respectively. I ran the numbers again and again just to make sure I wasn’t dropping a zero (or adding one). But the result remains the same – the fastest and most modern supercomputers are much less per MFLOP than their predecessors.

So what have we learned today? First, even though the cost of computing components has dropped radically over time, the cost of building/buying the biggest landmark computers in the world has increased just as radically.

We’ve also seen that today’s home and business computers stack up very well against supercomputers of the past. The iPad 2 pwns the Cray 2 in Linpack, but costs 99.998 per cent less (leather case not included).

We’ve learned that the current chart-topping supercomputer, the K computer, has a competitive cost per MFLOP/s even though it cost something like $1.25bn to design and build.

But my wife’s desktop is the Linpack price/performance king, and now we know that too. ®

This article originally appeared in The Register. It appears here in its entirety as part of a cross-publishing agreement.

SeaMicro is the server upstart founded four and a half years ago that came out of stealth in June 2010 with an Atom bomb: a 10U chassis crammed with 512 single-core Atom servers in a 10U chassis, all linked by a 1.28Tb/sec 3D torus interconnect implemented on an ASIC that also virtualized all I/O in the cluster of server nodes and provided load balancing across those nodes.

The acquisition of SeaMicro will no doubt irk Intel, which has used SeaMicro as a poster child for its microserver and cloud computing efforts in the past two years. So there’s some public relations fun in that. But there isn’t much Intel can do about SeaMicro being in AMD’s hands.

In a conference call with Wall Street analysts announcing the deal after the market closed, Andrew Feldman, SeaMicro’s CEO and co-founder, said that his company had been buying Atom processors and chipsets from the distribution channel, not from Intel. SeaMicro does not anticipate any changes in getting its supply of Atom and Xeon processors or building systems for current and future customers.

In the call, Read reiterated that AMD was “seizing the inflection point” in the server business. The acquisition of SeaMicro, which along with Cisco Systems is one of the first new entrants in the server racket in years to actually look like it has a chance of making it, would “accelerate out transformation into an agile and disruptive innovator.”

As we can now see, looking at the past month in hindsight, is that AMD figured out that is needed to hunker down on its Opteron processor roadmaps and not force server makers to absorb a socket change in 2012, as was the original plan, but rather to get an improved “Piledriver” core for Opteron processors into the field this year and do lots of performance tuning to show what “Bulldozer” class of chip can really do. The company then looked around and decided that it didn’t need a new chip or chipset so much as a whole different set of technologies to sell to its server OEM partners. That’s where SeaMicro came in.

Much will be made of AMD entering the server business with the acquisition of SeaMicro. But that’s not the point, even if it will be technically true over the next months. Lisa Su, senior vice president and general manager of the new Global Business Units at AMD, who is tasked with making sure the company has the components from which server makers want to build systems, put the kibosh on that idea.

“This is a technology play for us,” Su explained, calling the interconnect used at the heart of the SeaMicro system a “unique technology” that would be a “building block” that AMD integrates with its processors and chipsets and then in turn sells to server makers. The company will work now to get AMD’s Opteron server processors embedded in the SeaMicro machines, which it expected to have accomplished by the second half of this year. The company will not divulge what Opteron chips it will weld onto the SeaMicro server boards, but it stands to reason that it would be a low-voltage Opteron 3200 or 4200 part.

In an interview with El Reg, Mark Papermaster, AMD’s CTO, deflected all of the interesting questions about how the SeaMicro 3D torus interconnect might be used, such as extending it beyond one chassis to span multiple chassis under a single management and cluster domain, or maybe adding cache coherence and shared main memory across all or a portion of a SeaMicro server’s nodes.

The acquisition, said Papermaster, “allows us to redefine the server building block,” and he noted that AMD was keen on becoming “an agile SoC company,” referring to system-on-chip designs that put processors and other system elements into a single piece of glass.

And he confirmed El Reg‘s sense that AMD was not getting into the server business. “You have that exactly right,” said Papermaster.

When asked about when AMD and SeaMicro started talking, Papermaster would not be specific, except to say that AMD “moved very, very rapidly.”

This was quite possibly to keep someone else from buying SeaMicro and very certainly in part as a reaction to Intel’s purchase of the InfiniBand networking business from QLogicback in January for $125m.

Papermaster would not comment on rumors that AMD has been taking a look at ARM server chip maker Calxeda, but it could very well turn out that if AMD had tried to do that deal, it might be grateful that it fell through. The SeaMicro deal is a much better option if it wants to sell server interconnects and help OEM customers build cloudy servers and maybe even clustered supercomputers based on the SeaMicro “Freedom” fabric, as the latest generation is called.

That fabric is processor and instruction set agnostic, as Feldman has said many times, and SeaMicro only started out on Atom processors because they gave the best bang for the watt. And it could turn out that Calxeda and AMD partner to make ARM servers somewhere down the road. Why not?

Read wanted to make sure that everyone understood that the SeaMicro acquisition would have no effect on the company’s 2012 and 2013 Opteron processor roadmaps and system plans, and that AMD was not leaving the traditional x86 server business. “We’re in that business to stay, and we’re going to grow it and invest in it,” Read said.

AMD is paying $281m in cash and $53m in stock to acquire privately held SeaMicro, which raised $35m in two rounds of funding and also got $9.3m in funding from the US government as part of the $787bn American Recovery and Reinvestment Act (ARRA) of 2009.

Feldman is staying on at AMD after the acquisition closes and will be general manager of a new Data Center Server Solutions business unit, reporting to Su. ®

This article originally appeared in The Register. It appears here in its entirety as part of a cross-publishing agreement.