Slashdot Mirror
Tera Completes Acquistion of Cray
dewey writes "Tera, a new kid on the supercomputing block, has successfully completed its acquisition of the supercomputing pioneer, Cray (formerly owned by SGI). The new company will take Cray's name.
Tera has a press release from a month ago that spells out some of the details of the deal.
"
When I was job hunting in the early eighties,
the corporate status symbol was to have a Cray
super-computer. These were polygonal towers
(good shape for minimizing slow wire lengths)
with a encircling bench at the base. The job
interviewers would invite you to sit on the bench
as the ultimate Geek perk.
A decade later, on a trip to China, companies
would show me their Galaxy supercomputers,
identical in shape to the Cray-1, but a little
larger. (Probably identical circuits too.)
They've been around since the mid-80s,
"old-age" in the computer industry.
Someone there has deep pockets,
beacuse they've only shipped a couple of
machines for revenue according to their reports.
Here's hoping Tera/Cray at least re-evaluate some of his designs.
Apparently, a lot of people here like to shout "beowulf cluster" and don't understand much about the problems of massively parallel computing.
The bottleneck is inter-processor communication. If all you are doing is trying to brute force a cipher the processors are almost independent and easily reach the theoretical aggregate performance figures. But if you are doing complex physical simulations you can end up waiting for data most of the time and using only a fraction of your theoretical parallel power. Well spoken!
We've just been through an evaluation cycle where we pitted all major HPC systems and most Intel/Linux clustering solutions against each other. While I can't post numbers (yet), the results are that Intel/linux clusters come out at the bottom, mostly because of high latencies in the interconnect. I think it is not only the hardware per se, but immature drivers also.
We looked at Myrinet, Giganet and SCI. All of them were pretty close in performance, but don't come near more established SMP, NUMA or clustering technologies under traditional Unices. Our benchmark is molecular dynamics code with a reasonably large solvated protein.
Announcing the 2002 Cray Research model T3r4. If the name doesn't make you laugh, you should stop reading now. You have an obvious inability to truly appreciate the power of our newest offering. Go find someone who laughs and ask him to hit you with a clue-by-four. Good! The new T3r4 comes in three standard massivly para.....
.sig: Now legally binding!
And once more: Big Iron still has its place. It's not anymore as prevalent as it used to be, but standard PC architecture has design limits that prevent it from ever being usable for certain tasks (basically, anything that requires low-latency, high-bandwidth I/O). The supercomputing center next door just got a new Big Iron that's now the most powerful computer in Europe. It cost about $30,000,000. Look at the specs, especially the IO part. A Beowulf for the same money would be a huge, mostly useless pile of junk, when faced with the kind of problems this machine is designed to solve.
The illegal we do immediately. The unconstitutional takes a little longer.
--Henry Kissinger
Spoken well as someone obviously unfamiliar with the technology. IBM is aiming for a new peta-flop box - how many P-III/1GHz machines do you want to string together to:
On top of that, what are you going to connect the nodes together with? I don't think gigabit ethernet is going to handle the load you're putting on it. Ultra-high-speed bus handling is part of the magic that makes supercomputers truly super.
Beowulf is neat, and I'd love to experiment with it myself. It can even potentially replace a mainframe for transaction processing, although what OS you'd run for the task is beyond me. However, supercomputers are in their own unique class, and a few (million) Linux boxen aren't going to be able to compete in power, cost, usability, maintainability, or any other arena that requires the true Big Iron.
Dewey, what part of this looks like authorities should be involved?
This discussion is confusing two important aspects of supercomputing: the viability of the supercomputing concept, and the viability of supercomputer companies.
As a concept, supercomputing is and will remain tremendously viable and important - we'll always (well, OK, for the foreseeable future) be able to engineer fatter, lower latency interconnects within a box than between them as some sort of network. Still, the network guys are closing the gap some, so much so that many former supercomputer customers are now using clusters of some sort or the other to replace supercomputers.
As a business proposition, supercomputing is suicidal. The tidal wave of Moore's law and decently performing commodity hardware and software is firmly against you, not to mention these things take a LOT of man-years to develop and sort out, making it much harder to present customers with a value proposition that makes them willing to part with megabucks. Years ago, it was true that custom silicon gave you a real advantage at the high end - but no longer. The bottom line is that lower cost solutions are becoming "good enough" for the hard problems most people are interested in, so the market is shrinking even as development and manufacturing costs soar. (Remember how much of the economic viability of technology is determined by volume, something these guys will never have, almost by definition.)
Cray has been a walking corpse for years (seriously, go back and see when was the last year they made money - I'll bet it was nearly a decade ago) - even most of their industrial customers have abandoned supercomputers, leaving only a few high end science projects and the spooks as their market. My guess is that if it weren't for the NSA et al, these guys would have been out of business 5-10 years ago.
"The future's good and the present is nothing to sneeze at." - Roblimo's last
I was one of the designers of the CM-5 supercomputer at Thinking Machines Co. back in the '80s. One of the terrible things about designing a supercomputer is that it takes so long, Moore's law has moved the underlying technology out from under you before you're done. We had to throw away a mostly-completed microprocessor design when it became clear that we would be beaten by commodity microprocessors, and re-engineer the rest of the system to use a big pile of commodity SPARC chips instead of a big pile of our custom whiz-bangs. It was an enormous waste and really bad for morale (mine, anyway). So we ended up slipping one generation.
I find it hard to imagine what it must be like to work at Tera. I read Burton Smith's first paper, proposing what has now become the MTA, in 1985. Computing has gotten a thousand times faster since then, and they're still plugging away. Imagine the number of designs they have had to throw away! The've probably had to redesign the system every 18 months, just to stay even, without ever being able to build it, until they managed to deploy one machine, last July. It's an incredibly long slog, espescially for an industry like computers, that rewards sprinters, not sloggers.
Cray --> SGI/CRAY --> SGI --> Tera --> Cray
Anyone else see a vicious cycle?
Whaddya think they're goinbg to use for a slogan:
Cray: As cool as we used to be?
Cray: Not your mothers Supercomputer?
Tera: We put the T in T3E?
Cray: 100% SGI-free?
Cray: More expensive than a failed Pesidential campaign. (starring John McCain)?
Cray: We were chewing numbers when Beowulf was just a gleam in some cheapskates eye?
Cray: Buy Big Blue? We've got chunks of IBM in our stool!?
Cray: Stop drooling and buy one?
Cray: Just plain sexy.
.sig: Now legally binding!
Apparently, a lot of people here like to shout "beowulf cluster" and don't understand much about the problems of massively parallel computing.
The bottleneck is inter-processor communication. If all you are doing is trying to brute force a cipher the processors are almost independent and easily reach the theoretical aggregate performance figures. But if you are doing complex physical simulations you can end up waiting for data most of the time and using only a fraction of your theoretical parallel power.
This is especially true of things that don't break down nicely into regular blocks. For a wind tunnel simulation you can assign different sections of the tunnel to different processors and each one communicates only with its neighbors with nice, predictable access patters. For something like a simulation of a car crash this will not work too well, though.
Tera's architecture is based around a high throughput communication fabric. The cost of this architecture is the latency - it can take many cycles from the time you request a piece of data stored in another processor until it traverses the switch fabric and the result comes back. To get around this problem each processor runs many threads with very fine granularity - it switches to a different thread every instruction cycle. By the time the next instruction for the same thread is scheduled for execution the results of a remote memory access are already available, without wait states. Each of these "virtual processor" threads is not particularly fast but the total throughput is very high.
This presents the programmer with a simple shared-memory multithreaded programming model. No need to reengineer your program to a specific message passing architecture supported by the target machine.
----
Stop worrying about the risks of nuclear power and start worrying about the risks of not using nuclear power.