Cyclone will offer Software as a Service (SaaS) with third-party applications in five areas of "functional computing": computational biology, computational chemistry and material, computational fluid dynamics, finite element analysis and ontologies (data mining). The service will also let users program directly to SGI's infrastructure platforms. Analysts called the move a realistic stab at appropriating earlier attempts by Sun and IBM to sell HPC on-demand.
"It's an early adopter's market, but it's real," said Steve Conway, research vice president for IDC's HPC group.
He said that while IDC projected cloud computing -- pay-per-use, on-demand consumption of IT -- to be around 9% of total IT spending by 2012, that's still around $42 billion. He added that interest in general purpose GPU computing (GPGPU) was particularly high among traditional consumers of utility computing, for obvious reasons.
"Among oil and gas companies, the majors are definitely looking at this," he said. "They've been using utility for years and years." They want to see the high overhead and long-term contracts from utility computing disappear in favor of more flexible pay-as-you-go services.
Conway said they'd already hashed out the problems around data transfer and latency to operate with utility computing services, and they were also forerunners in wanting to use GPUs for specialized, compute and graphics intensive projects.
SGI said that it is aiming for its first customers to be "bridge customers" that want to get results faster while building or upgrading their own capacity, along with customers already using one of the software services provided for extra capacity.
Taking over where Sun left off
This, however, has been tried before. It's a close match to Sun's Compute Cloud service that foundered mid-decade, despite a similar pitch and platform. It's even close in price, something that is not in SGI's favor, said Conway.
"There were buyers, but [Sun Cloud] didn't really take off because a dollar per hour was a very high bar," Conway said.
IDC projections put an attractive price for this kind of infrastructure service at around $0.20-$0.25 per hour, two thirds less than SGI's current $0.95. But SGI isn't betting the farm on this, and it has plenty of expertise in distributed computing and the cloud. Rackable sold servers to Google, Amazon, Microsoft and many others doing cloud, and SGI brings decades of expertise in symmetric multiprocessing (SMP) computing.
Evaluating SGI's efforts
"Honestly, the way I see this, it's SGI trying to get out there early with experimental capabilities to see what the market interest is," said Frank Gillett, VP and principal analyst at Forrester Research.
He said SGI sees an opportunity to capitalize on two years of "breathless cloud hype" and succeed where Sun failed at the same price.
"Some time has passed and now there's a much wider education of the idea that I can rent capacity from somebody, so there's more education, more comfort with it," he said. Gillett said that SGI was tackling a very specific subset of workloads that it could directly tie to business units and researchers instead of getting lost in the enterprise IT shuffle.
Gillett called it a plausible offering but said that the GPU platform was far and away the most interesting part of the announcement. However, despite unbeatable supercomputing credentials (Cyclone is built in part on SGI's Nehalem Altrix UV line, reportedly the fastest supercomputer in the world by some benchmarks), SGI faces competition.
Utility providers like 3Tera notwithstanding, SGI will need to pitch against up-to-date companies like Penguin Computing, which has a much more flexible HPC cloud service and announced GPU compute services last fall, and ZetaExpress, which has a firm grip of oil and gas users as well as academia in GPU Infrastructure as a Service (IaaS).
SGI said it wants an initial commitment to buy at least one week (144 hours) of computational time, and it's definitely not self-service. SGI will put buyers in touch with an engineer after a sale is made to draft a statement of work before beginning a computing job. That's a cold-water bath to users charmed by cloud computing's DIY approach.
The service will run in two data centers, one in the Silicon Valley and one in the Midwest. The company declined to say how many cores it was stocking, but of, course; it can always add more if demand goes up.
Carl Brooks is the Technology Writer at SearchCloudComputing.com. Contact him at firstname.lastname@example.org.