I wrote about the demise of SiCortex when it happened, but the subject has surfaced again recently – first in a Network World article, and then on Google’s Cloud Computing group. I’d like to address a few issues these raise, plus some that have come up elsewhere. Please bear in mind that these are all from my perspective as an engineer in SiCortex’s software group, so more informed parties might be able to correct some of the details, but I’m pretty confident about the gist of them.

  • SiCortex was explicitly not trying to sell a supercomputer. We never even cracked the Top500, or intended to except as a prerequisite for the more relevant Green500 (at the time; the requirements have since changed). We even offered the 12-node 72-core $15K SC072 as an entry point, which was lauded for its affordability and accessibility (and predated Cray’s CX-1 by quite a bit). That’s hardly “Formula One” territory.
  • The “green” angle was not just a matter of market positioning, though it was certainly that. Low power and heat were also quite fundamental to achieving the level of density and simplicity (e.g. no active network components other than the compute nodes) that were the real core of the architecture.
  • Using an x86 instruction set was not an option. Even if the appropriate licenses had been available, they would surely have been out of our financial reach. As it was, IP licensing (especially MIPS) was the second largest expense behind payroll.
  • Similarly, using QPI or similar wasn’t an option. QPI itself didn’t even exist at the outset, and licensing it later would have been similarly expensive.
  • Both the x86 and QPI mirages are also affected by the question of technological fit with other key architectural elements (see above). It seems vaguely possible that something could have been worked out with VIA or some other third party that already had the necessary licenses, but then again maybe not. The same question came up with respect to Raza and Cavium in the MIPS space. I was personally never very satisfied with the answers I got when I or others asked about such things, but then I was barely qualified to be asking. Maybe one of my erstwhile colleagues could shed more light here.
  • It was probably not feasible to make and sell machines even smaller than the SC072. I and others did favor this idea, largely based on the four-node “Frost” machines we’d used during development, but the people making such decisions said that we could not afford the sales and support infrastructure necessary for that market. Having seen the margins and support costs for the SC072 up close and personal, I agree that it would not have been profitable. An argument might still be made that such a machine would have been worthwhile even if it had not been profitable in and of itself, as a way of establishing mind share etc., but that’s not clearly the case either.

As I said in my other post, the failure of SiCortex had nothing to do with the technical merit of the product. I don’t think they even had to do with the market merit. Don’t believe the disgruntled ex-employee who now works for another player in this space and seems to make a point of going around to every article or story about SiCortex to comment anonymously about CPU or memory performance. Sour grapes, dude. The machine was selling OK, though of course not as well as anyone had hoped. Where we failed was in “selling” the company to investors so we could finish building the next generation, and in any other VC market even that might have been an easy sell. After all of the times I’ve seen charlatans do a better job of selling a story to investors than of selling a product to users, I think SiCortex still stands as an example of the exact opposite. We were for real. It’s the investors who turned out to be fake.