Every once in a while I get credible seeming "deepthroat" e-mail -unfortunately these aren't always completely right, as witness this report last January from an Apple insider saying that the iPhone's application processor would be a PPC enabled core made by Samsung. Although he may have been completely right at the time, he appears to have been only half right for the production gear - and that same kind of uncertainty plagues last week's winner in this category: telling me that Sun is working on a Niagara variant aimed entirely at short term release into the dedicated HPC market.
I have no basis for judging this information - although the floating point process improvements they're about to announce for the second generation Niagaras combine with the latest bladeserver announcements and the new superswitch to suggest that some such scenario might make sense.
In thinking about this, however, I came up with a fascinating speculation about how this might relate to IBM. You may not have noticed, but there's a Texas sized silence emanating from Armonk about progress getting the Cell2 software to work.
There's a stage magic aspect to this: highly visible actions intended to get you to watch the right hand while the left makes the switch - or, in this case, Power6 and AIX6 announcements aimed at helping the IBM faithful maintain the right while distracting the audience from a lack of announcements on the second generation cell processors.
Thus IBM recently announced an AIX6 beta - but the product seems to be the same old System V stuff with IBM enhancements like mealy mouthed editor messages, data processing oriented management "functionality," and lots of largely unintegrated Solaris 10 features. Similarly they've announced, and I believe will soon be shipping, the 4.7Ghz Power6 - a huge gigahertz jump unaccompanied by commensurate bandwidth gains and therefore something that doesn't actually do much for real world system performance.
John Meyers has been using his Sun blog to comment extensively on on IBM"s Power6 performance problems - which he describes as the Power6 Thud. Now I'd be surprised if he knows how The Thud earned its nickname, but either way his analysis includes lots of useful benchmark information the bottom line on which is that the the Power6 beats the power5+ by 114% on gigahertz but only 58% on over simplified benchmarks like the TPC/C -and generally does significantly worse on more complex jobs.
The reason for this is simple: IBM is running out the clock on the Power series - and what you should see in both the Power6 and AIX announcements is IBM's right hand waving speeds, feeds, and features at the public while the left hand gets ready to switch everything they make to Cell.
Thus the explanation for the halfway house design running the Power6 at half the performance per cycle achieved by its predecessors is that this is really an exercise in manufacturing and software reconaisance by customer trial, with the new stuff either designed for Cell or asserted to sell - thus the high cycle rate tests manufacturing, the instruction set changes are intended to help work around binary compatibility issues with Cell's master processor, and everything they're saying about AIX6 is adaptive, check box, stuff designed only to sell against Solaris.
And that brings us back to last week's voice from the underground: what is Sun planning, if anything? People are still buying x86 based grid computers - but really only because of organisational conservatism and multi-year budgeting cycles - Cell, and particularly cell2 with its enhanced double precision capabilities, owns the market for anything entering the planning stages now. IBM, however, seems to be meeting delays on the software side: and a spoiling play by Sun, if they can get the technology out soon enough, could disrupt some very serious strategic execution work going on in Armonk - meaning that if they're not actually doing this, I'd be surprised if they aren't talking about doing it.