Cray Co-Founder Joins Microsoft

ergo98 writes "Burton Smith, co-founder and chief scientist at Cray (The Supercomputer Company), has jumped ship. He's joining Microsoft to help them with their clustered computer initiative. Burton joins Microsoft as a technical fellow."

Re:Irresistable

[in7ane]

They were bought by SGI in 1996, then spun off and sold to somebody else, who then renamed themselves Cray once again - so Cray is indeed the supercomputer business. Somewhere along the way their not-so-super computer business was sold off to someone else. And no, it is more than name and reputation, they sell the Cray X1 and had some clustering product coming out, which could be hurt by this departure I guess.

And that's some very interesting logic - if you are not no1, just give up.

Mr. Proprietary, meet Mr. Proprietary...

[Nigel_Powers]

Isn't Cray hardware and software completely proprietary? If so, no wonder MS is interested in teaming up with Burton Smith. However, as this article [slashdot.org] suggests, Linux is way ahead of the curve in this arena.

Linux may not ever truly catch on in the desktop environment, but in high-end computing, it's a proven winner.

Re:Irresistable

[geoffrobinson]

With all their cash, they can catch up in a big hurry. Also, with their market position they can bide their time. How long did it take NT/2000/XP to become somewhat respectable?

Re:Irresistable

[Frumious Wombat]

I think that "not so super" product you're referring to was the Sparc-based system, which became the Starfire E10K. SGI/Cray couldn't make money on it, but Sun used it to eat their lunch.

Like the old IBM, Microsoft is now big enough that various pieces are running their own projects, and it will be interesting to see how this plays out. Windows that seamlessly clusters, where you could just add machines transparently in a manner similar to a Condor flock, would be an interesting competitor. They may be a lumbering, http://www.eps.mcgill.ca/jargon/jargon.html#evil%2 0and%20rude/ [mcgill.ca] Evil and Rude corporation, but there are some really bright people in there working on more than Office.

Inevitable, really...

[Anonymous Coward]

I think it was a foregone conclusion that Smith would eventually leave Tera^H^H^H^HCray once they dropped the MTA as a product.

Some people have acted as if Burton Smith is the second coming of Seymour Cray. To be blunt, I just don't see it. The MTA was Smith's baby, and by most accounts it was a failure. The first version of machine was based on gallium arsenide technology and was very problematic to manufacture; less than 5 were built. Tera bought Cray largely for their CMOS design experience because they wanted to convert the MTA from GaAs to CMOS, but even that wasn't enough to fix its performance problems. While the massive multithreading capability is cute in theory, the MTA architecture simply doesn't have enough memory bandwidth to handle the scientific codes that cause people to spend 7-8 figures on a supercomputer.

It does seem weird that Burton would go to a software company like Microsoft, though. OTOH, Microsoft Research also employs Jim Gray and Gordon Bell...

Unix

[Dollyknot]

Perhaps he is going to teach Billy boy Unix, the defacto clustering OS. It could be said, the internet is Unix based, Google is Unix, Apple is Unix, Amazon is Unix, I could go on - Beowolf anyone? 'Tiz a shame Billy boy did not complete his computer science education.

Re:Irresistable

[Analog Squirrel]

I think the "current" incarnation of Cray started when they were bought by Tera Computing, whose primary contributions to supercomputing are in massivly multi-threaded computing. Not the wimpy hyperthread that intel has - 128 complete sets of registers per processing unit, data/control flow analysing compilers to automate the extraction of threads from a program, and a huge, proprietary flat (no cache) memory architecture to make sure that the processor always has instructions and data to compute with. I remember seeing Tera at the Supercomputing 1999 conference... and they've likely improved since then.

Re:Clarification of "co-founder"

[mpg]

Burton Smith responsible for architecture of the Tera MTA series and, much earlier, the Denelcor HEP -- both of which were ahead of their times technically but complete failures commercially. (Indeed, Tera Computer had significant financial problems and some corporate governance issues in the years leading up to the Cray purchase. I don't know the financials of Cray today, however.)

Some thoughts, in no particular order:

* The MTA and the HEP, together with Multiflow, represent the commercial roots of the multithreading (MT) work still going on in academia today. Note, however, that the "real" MT work is different by an order of magnitude from what we see in the threaded commericial chips emerging now from Intel, etc.

* The rumor as of a year or so ago was that Burton and a few of the Tera old guard had been pretty much sidelined from the larger Cray operation into unfunded R&D projects being pitched to organizations like ARPA, etc. It would be nice to believe that someone in the commercial arena is going to fund traditional MT ideals, but I'm skpetical.

* What is Microsft doing hiring him? Is this a largely PR move, to improve their HPC image? I have a hard time believing Microsoft is going to spend any money doing parallel architecture work; the list of companies that have tried and failed is long and impressive. Supercomputing today is either custom stuff, or high-end-but-nonetheless-stock hardware running Linux clusters. What's their angle?

* Back in the day, Tera had one of the hottest compilers on the planet; indeed, their compiler IP was pretty much the only valuable stuff left from the MTA project. [Ditto for Multiflow, whose compiler served as the base for Intel's compiler, way back when.] It would be interesting to see who else from the original Tera team follows him over to Redmond -- compiler folk? Architecture folk? Surely not hardware folk?

* If Microsoft wanted Burton, did Google make a play for him too? Now that would have been interesting -- one could have a fun time speculating about masive parallelism and large-grained work tasks across Google's distributed network...

[disclaimer: I briefly worked at Tera in the late 90's.]

Re:There's a difference between megahertz...

[forkazoo]

Well, the Cray-2 was released in 1985, so it only just barely qualifies for the OP's criteria, since it would not have instantly been in widespread use. The X-MP Was released in 1982, so it is probably a fairly good guide to what would have been a "common" supercomputer at the start of 1985.

Before the XMP/EA's came around, the XMP had a max memory capacity of 128 MB (stated at the time as 16 Megawords, as byte notation was not yet universal.) 4 Processors, and a theoretical peak of 200 MFLOPS per processor. Thus, about 800 MFLOPS theoretical aggregate peak.

I just looked up a few numbers real quick... Looks like a dual-proc, dual-core Opteron 270HE has a theoretical peak of over 17 GFLOPS. I'm not intimately familiar with the memory latency characteristics of a cray, but I really can't imagine there being much competition between the two, no matter how great the IO was in 1985.

Obviously, quad-core Opterons are fairly high end... dividing out, and a single core from the system I was looking at the numbers for would be about 4 GFLOPS. Of course, that's peak. Probably something like 2 GFLOPS easily sustained for a modern single desktop CPU. Any AthlonX2 should be able to run the old nuclear sim code quite a lot faster than the "average" cray at the start of 1985. Regardless of any verbal mis-steps, or name calling in this thread, I think the original point was well made. I'd love to get to play around with some of the old sim software. Let's break out the g77, bitches! Let's get a nuclear sim project on sourceforge. It'd be greatly educational, both from a retrocomputing perspective, and from a physics one.

High-speed interconnects

[jd]

These days, a "high speed interconnect" means doing Infiniband better. Many of the exhibits at the SC2005 show were using Linux, OpenIB and Inifiniband, which is a good start - but slow, because Infiniband is generally implemented as a pseudo-bus run on top of PCI or PCI Express. The added layering adds a lot of latency, and it is latency that is killing a lot of high-end applications. That, and the fact that fat-trees saturate so easily, killing performance.

Further research into the NEC/AT&T connection

[einhverfr]

WEC (the manufacturing arm of AT&T) joined with Japanese investors to create NEC in exchange for a 54% stake in the company. So while the companies were legally separate, Ma Bell essentially controlled NEC as if it were a Japanese subsidiary of WEC. So while NEC might not be a baby bell in the classical sense, it is certainly to be argued that it was effectively such a company.

It never ceases to amaze me how big Ma Bell was.

They Need Something - Unisys Didn't Help

[Anonymous Coward]

Microsoft tried to partner with Unisys on multiple CPU architecture but, unlike other partherships, Unisys screwed M$ instead of the other way around. Gotta give credit to those sleasy Univac bastards! Old age and experience trumps eagerness and youthful vision every time.

Microsoft needs someone or something to get their multiple CPU and clustering architecture working halfway well. They don't seem able to do it themselves. But I predict this effort will fail too.

Branded!

[fm6]

After their big purchase, they took the Cray name for continuity with Cray's old customers and products, along with the fact that it's a much more viable "commercial" supercomputing name.

"Continuity" is kind of the wrong word, since the SGI had little use for the Cray name. While they owned Cray, the name appeared only on minor products such as the Craylink bus.

Despite SGI's neglect, the Cray name did and does have a lot of name recognition. So when Tera bought SGI's Cray division, they did so not just for the right to restore the Cray name on Cray products, but for the right to put the Cray name on Tera products. It's an exercise in branding [wikipedia.org]. Indeed I suspect that Tera was more interested in buying the Cray brand than the Cray product line — which has never been profitable.

A more extreme case of branding is Atari, which is now the name of a French game software company [infogrames.com] that has no real connection with Nolan Bushnell's original company.