About Dropouts, Disclosures and Delays

Andreas Stiller

(Translation of the German original in c't by Marcel Sieslack)

First, IBM celebrates its 100th anniversary, 50 years of IBM Golfball Typewriter and 30 years of IBM PC – and then it has to silently drop the prestige project Blue Waters. But the other players in the game are also merrily postponing their projects.

It came as quite a surprise when IBM and the National Center for Supercomputing Applications (NCSA) in Illinois announced the cancellation of the planned supercomputer project Blue Waters. That the current U.S. economy crisis is to blame is unlikely – at least, 3.7 billion dollars quarter profit should have been enough to compensate for possible financial constraints on the side of the project's sponser, the National Science Foundation.

But, apparently, the costs for the ambitious 10 petaflops supercomputer project with 320.000 Power7 processor cores were completely out of control; as is the case with certain philharmonics, opera houses or – soon – underground train stations elsewhere (a very disputed project in Stuttgart, Germany) . The contract with the NCSA, signed in 2007, raised 208 million dollars, but such peanuts don't suffice to bring a project of this scale to fruition. Unexpected technical difficulties are cited here and there.

What's worse, IBM had already lost the prestige duel against Fujitsu in the run-up – the Japanese just did too well with the K Computer. Since June, it spearheads the Top500 list of supercomputers with 8.2 petaflops at only 85 percent of completion. So Blue Waters would probably never take first place. And maybe it was all just a dispute with the NCSA because they weren't willing to run a Linpack benchmark.

Blue Waters Run Aground

The start date, originally aimed at 2011, had already been pushed back to mid-2012, anyway. Considering the not otherwise specified technical problems, it would most likely even have been significantly later. But while opera houses à la Sydney are appreciated in spite of an 8-year delay, supercomputers that miss their narrow time window are obsolete before they go on line.

Blue Water's impressive "motherboard" advertises "Cost Effective Quality Design", but apparently it wasn't so cost effective, after all ...

In any case, IBM wasn't making a profit with its "Deep Computing" branch. This high-end sector of computer technology primarily served prestige purposes. This had been a thorn in the side of IBM's controllers for some time already and so, in November last year, Herb Schultz, marketing manager of IBM's Deep Computing Group, had announced a reorganization of the HPC business toward the more profitable midrange sector.

Still, IBM doesn't intend to quit the Formula 1 completely – after all, it has another promising racer on the track. It's not as spectacular but it's causing less difficulties, besides being cheaper and faster: the Sequoia, for the Lawrence Livermore National Laboratory, with its 1.6 million BlueGene/Q cores is supposed to deliver more than 20 petaflops next year.

Just as once Stauber (from BMW to Ferrari) or Williams (from BMW to Toyota and Cosworth), the NCSA racing team has to look for a new partner now, because a mutual agreement for an alternative project could not be reached. Cray and SGI are being considered as promising candidates. For reasons of state, new partners probably have to be American – if otherwise, contacting Fujitsu might be a good idea ...

IBM will now return the received payments and the NSCA will have to cough up the IBM hardware that had already been delivered.

If IBM's reorientation will effect the supercomputer SUPERMUC, ordered by the Leibniz Computing Centre, is still unclear. It's based on very different hardware: iData-Plex with Intel processors. Delays are possible for other reasons, though, as rumors are merrily pushing the release date for the Sandy Bridge-E and -EP processors with six and eight cores back and forth on the timeline. Latest news: the desktop versions will arrive at the end of the year, the dual-processor versions for servers only during the first quarter of 2012.

Other Delays

AMD, it's said, plans to launch its Bulldozer processors for servers on the 26th of September, one week after releasing the FX versions. That's later than initially planned, but still early enough for a headstart of a few months on the competition. After all, these heavy machines don't only come with up to 16 integer cores, but also are the first AVX capable, real server processors for more than one socket. It's somewhat amusing that the copycat is earlier than the original here.

Nvidia's Kepler and Intel's Sandy Bridge-E seem to have a common motto: to and fro. I still believe that Kepler will, completely coincidentally, be released at just the right time for the GPU Technology Conference, which has been rescheduled to take place in May 2012.

Also the Chinese Loongson or Godson 2H processor will arrive significantly later than anticipated. According to semiaccurate.com, this was confirmed by professor Yunji Chen from the Institute of Computing Technology (ICT) of the Chinese Academy of Sciences. In any case, the Chinese processor designers' first system on a chip (SoC), including the memory controller and a graphics processor licensed from Vivante, still is in its tapeout phase and thus won't be around any time soon. At least, all legal animosities with MIPS are dealt with; actually, an agreement had been reached last year already, and now Loongson Technology officially received the architecture licenses for MIPS32 and MIPS64.

At the Hot Chips Conference at Stanford University in mid-August, the Godson 3C and its impressive features will probably be at the center of attention again. However, until 2013, it remains a paper tiger. First, the Godson 3B, currently being manufactured in 65 nm technology by STMicroelectronics, would have to materialize in adequate numbers to push the supercomputer Dawn 6000 into the petaflop domain. It had been expected last year already, but then it was announced that a smaller pre-version with 300 teraflops would be presented this summer – which means the Chinese doesn't have a lot of time on their hands. Until they will be able to compete with Intel with proprietary processors, it's still a long march. Head architect professor Hu even talks about 20 years.

Oh, and I almost forgot: At Hot Chips, Intel intends to shed some more light on the next Itanium processor Poulson. So maybe, there will soon be some interesting hardware, but almost no up-to-date software, as, amongst others, Microsoft, Red Hat and Oracle – the latter very much to the displeasure of HP – ceased any further development of their software for the Itanium.