Original URL: https://www.theregister.com/2009/04/01/rackable_sgi_supers/
SGI's Rackable's future supercomputers
Goodbye Itanium, hello Nehalem
Posted in Channel, 1st April 2009 22:06 GMT
Comment The Chapter 11 bankruptcy filing by beleaguered supercomputer-maker Silicon Graphics and the acquisition of most of SGI's assets by Rackable Systems for $25m in cash may lay to rest any questions about the future of shared-memory, Itanium-based supercomputers.
As Dell and IBM have encroached on Rackable's business among hyperscale Internet companies with their custom-built servers, Rackable has tried to peddle its energy-efficient and compute-dense rack-based clusters to supercomputer centers.
It can be argued that the $25m it's spending to acquire the assets and customer base of SGI is a lot smarter than the $40m in share-buybacks the company announced back in February when it reported fourth-quarter 2008 sales down 65.1 per cent to $38.8m and an $18.2m loss compared to a $4.7m profit in the final quarter of 2007.
Buying SGI is perhaps a way to kickstart HPC sales. It also seems to be a relatively inexpensive way for Rackable to get its hands on the intellectual property and expertise that will be needed to take a serious run at the HPC market - meaning commercial, academic, and government supercomputing centers that don't mind taking risks on new technologies if a vendor can cram the flops into a box.
Neither SGI nor Rackable returned calls for comment about the effect of the deal on the two companies' respective server and storage lineups - or, for that matter, on the employees at SGI -but a statement to customers from SGI's chief executive officer, Bo Ewald, gave some hints about the products.
"Our joint objective is to build an even more powerful company capable of providing a wide range of products from built-to-order clusters to our new breakthrough UltraViolet shared memory system that is currently in development," Ewald wrote. "We'll surround our HPC systems with combined storage and software products, and continue developing our leading edge VUE visualization software. We further plan to merge our two service organizations and continue to provide the customer and professional services for which you rely upon us."
UltraViolet, the code name for SGI's fifth-generation NUMAflex shared-memory clusters, was referred to as an "x86 shared memory system" in a similar statement from Doug Britt, SGI's senior vice president of global sales and service - so I guess we all know what SGI had planned for the future quad-core Tukwila Itaniums.
The scuttlebutt was that SGI would ditch Itanium for Nehalem, but there was always the possibility that the Altix 4700 shared-memory supers could be tweaked to support either chip if the QuickPath Interconnect were on both styles of system boards.
Britt did not say "Itanium-based" when he referred to UltraViolet. And you can't blame him.
The continual delays in the Itanium roadmap have made it very difficult for SGI to get traction. When it launched its Altix line of supers at the turn of the millennium, SGI staked its future on the Itanium and Linux combo. And by the summer of 2006 (as it was in an early bankruptcy protection) the company decided that it could not afford to do MIPS-chip and Irix Unix development any more, and killed the products off.
At that point, the company was tied to the Itanium roadmap, just as Intel was finally waking up to the 64-bit Opteron threat.
Whoops.
Since then, even with the excellent Altix 4700 architecture, sales have been hindered because the standard in the HPC space is, for many customers, an x64 server node.
The benefits of shared memory cannot overcome the incompatibility and expense of Itanium servers, and SGI got crushed in those pincers despite some big deals for Altix machines. And the Altix ICE blade-style clusters didn't have the shared-memory secret sauce enabled by NUMAflex, and therefore didn't give SGI enough of an edge against generic clusters.
Mysteries remain
Why SGI didn't launch NUMAflex for x64 servers is a mystery, but I suspect that SGI knew that QPI was coming for both Xeon and Itanium processors and was waiting for that to come to pass before making the switch. Such a switch was expected several years ago, but was pushed out by Intel.
Why SGI didn't realize it could be burned by Intel product delays - considering that it was an early supporter of Itanium - and therefore port NUMAflex to server nodes using Opteron processors from Advanced Micro Devices is another mystery.
I've continually badgered SGI's execs about this, and they've been steadfast in their devotion to Itanium - right up until the second the Altix ICE machines were launched. And they were quiet about their future Itanium support right up to today's third bankruptcy filing.
Although SGI announced it had two-socket Nehalem EP blades for its Altix ICE clusters for sale as part of the Xeon 5500 launch, it didn't say anything about marrying Nehalem processors with NUMAflex.
But that has to be the plan, since no other plan makes any damned sense at all. A cluster of Nehalem EP server nodes with 21TB of shared global memory is something that SGI should be able to sell.
In his statement, Ewald reminded everyone that Rackable has lots of cash and would reduce SGI's debts, some of which it will assume as part of the acquisition of assets that will follow the Chapter 11 resolution. As Rackable's fourth quarter of fiscal 2008 (ended January 3 of this year) came to a close, the company had $171.9m in cash. Mark Barrenechea, Rackable's CEO, said that the company would spend up to one-tenth of that cash on additional research, development, marketing, and sales to chase new markets, such as the HPC sector.
Considering that Rackable has canceled its share buybacks, that leaves something like $57m to play with, and the SGI cash price only eats up half of that cash.
But the deal is perhaps more complex than that.
In SGI's second fiscal quarter, which ended in December, the company had $36.1m in cash and equivalents and $157.4m in long-term debt. SGI had $526.5m in total liabilities and $390.5m in assets. I'm no bankruptcy lawyer, but the Chapter 11 filing seems aimed at lowering the amount of liabilities that Rackable assumes.
You can bet that Rackable wants to conserve all the cash it can - hence it didn't simply buy the company outright.
And while the deal is open for competitive bidding, it is hard to imagine anyone else coming in right now. IBM is busy with its Sun acquisition, and Sun is busy with IBM. Hewlett-Packard is already up to its ears in server architectures, has all that Convex goodness in its Integrity machines, and wants to sell big Integrity boxes to HPC customers who want lots of memory for apps to play in.
It's a pity that Integrities don't support Nehalem, eh? Whoops, again..
Other players are equally unlikely to bid. Dell doesn't think it needs anything more sophisticated than two-socket x64 boxes and an occasional four-socket machine. Fujitsu is distracted by its integration of Siemens' IT biz and its partnership with Sun. Hitachi and NEC, which know all about Itanium, seem to be aloof about HPC except in their home markets. Verari Systems, which most resembles Rackable among x64 server makers, did not go public as Rackable did a few years back and may not have the cash.
Looks like Rackable wins.
But what it wins is somewhat unclear. Both Rackable's and SGI's sales have been so choppy in the past several quarters - thanks in large part to the ongoing Meltdown - that it's hard to make any prediction about what the revenue stream and profit potential might be for a Rackable boosted by SGI's products and customer base.
A lot seems to be hanging on that UltraViolet supercomputer, which needed to be here yesterday.
Well, on Monday, actually - but it wasn't. ®