Slashdot Mirror
Tera Will Buy Cray Research
I just found this short news in C|Net which states that Tera Computer will buy Cray Research from SGI for an undisclosed amount of cash, stocks and notes (although the Wall Street Journal estimates the price that Tera pays is less then $100 million, which is a fraction compared to what SGI had to pay when they bought Cray - $740 million). Tera is going to change their company name to ... Cray.
I've been following SGI for quite a while now.. They began focusing on the x86 line as they figured the 64-bit Merced/Itanium/McKinney/[fill in the blank] would take over the marketplace sometime very soon for visual workstations.
On the high end market, they purchased Cray, and continued work on the Orion line of supercomputers.
Bad move buying Cray, bad move focusing on x86 64-bit imho. They had to get rid of Cray as it was becoming excess baggage. I may be wrong, but Intel is having trouble getting their 64-bit cpu higher than 400Mhz, while Alpha are laughing happily to the bank on their 21264 chipset, while working on their 21364 chipset (1000Mhz, 64-bit I believe). Sun continues working towards the UltraSparc-3 chipset (750Mhz 64-bit) and UltraSparc-5 chipset (1500Mhz 64-bit).
Perhaps SGI should stick to high end 32-bit x86 chips for their workstation line, and continue pumping out MIPS cpu's on the beautiful Origin line. I understand fazeing out IRIX, but really..
Where will it end? I love SGI products! Are they organized enough to provide me with a reliable long term gameplan, and be able to produce their servers in a timely manner?
EraseMe
When it comes down to it, this really isn't too surprising given what's going on in the market and what happened to CRI as it grew into middle age. The market forces are obvious: lots of computing problems that used to be in the exclusive domain of supercomputers have been assumed by much smaller computers. Budgets of large government agencies have been dramatically reduced and with them, a large customer base has disappeared. If you're trying to make money, this spells disaster.
Furthermore, like any organization, as they developed into more of a "sustained business" organization rather than a "let's build the fastest computer ever" organization, it became more and more difficult to innovate as rapidly as before (this is one of the reasons Seymour left). In the hey day of large government defense contracts, this was not a problem. However, as budgets dwindled they ran into significant difficulty penetrating new markets (a $30M machine is not an easy sell).
It's really too bad as they have done some really neat things. In some senses it was the ultimate geek environment. In the engineering tradeoffs, speed always wins (which is why the things are so darn expensive). They used different wires for memory reads and writes, high memory interleaving, the I/O subsystems are multiple computers tasked for nothing more than I/O, vector registers make large FP computation chains very fast, no virtual memory (you can't use what you ain't got). In most cases, if it was slow they threw H/W at it. And then there was materials research for cooling and lots of other cool stuff.
I don't know what Tera plans to do with them, but unless they have a good way to penetrate into the business markets (and there are opportunities), before long Cray will simply be a name for the history books. It's really too bad considering all the contributions they've made to the computing world.
Besides, Cray-type supercomputers are over. All modern big machines are collections of single chip microprocessors. They have to be; the speed of light is too slow for a gigahertz big-box CPU. You can't even afford the slowdown for off-chip connections. Designing a fast CPU chip isn't worth doing unless you can sell a lot of them, so supercomputers have to use commodity CPUs. All that's left is architecting big multiprocessors and clusters out of commodity CPUs, which is mostly a software problem.
SGI has other problems, the main one being lack of focus. "We're a graphics company". "We're a movie and game tool company." "No, now we're a server company". "We're an NT workstation company". "Forget that, now we're a Linux company". They keep trying to find some niche that isn't disappearing, and it's not working.
I have been following both Cray and Tera for many years now. I have been saddened watching the last of the supercomputer companies wither and die. Supercomputers were always so COOL, but for most things, they just aren't so "super" any more.
I have benchmarked several of my back end utilities on cray systems, and one of them on the early tera machine (the early compiler exploded on the others). None of the single processor runs were as fast as a pentium III, and this was quite some time ago.
Understand that this was often branchy and recursive code running with only 3D sized vectors, so it isn't the sweet spot for traditional supercomputers. If I was doing nothing but multiplying 1k by 1k matricies of doubles, even a five year old cray would kick the crap out of the latest athalon. Unfortunately, none of my code looks like that.
I even spent some time thinking how I could restructure calculations into a vectorizable form, which might make a cray J90 competative. I wanted to buy a Cray! Of course, this was silly. It took less effort to make the code SMP friendly, and the payoff was much larger.
We wound up with a 16 processor SGI origin-2000 system, which has been easy to develop for and predictable in performance. We just recently bought an 8 processor Xeon, which is actually faster than our old 16 processor SGI, but at exactly one tenth the price (the downside is that it is maxed out, while the SGI still has tons of growth potential).
I program all heavy workloads in a parallel fashion now as a matter of habit, but it is easy to overstate the benefits of parallel systems.
The common lunux advocate position of "beowulf makes supercomputers obsolete" isn't quite right. Even with code that is already written in a parallel manner, there is a large difference between developing for a shared memory system and a distributed cluster. Developing a compute (and especially data) intensive program for a cluster rather sucks.
If there was a single processor system that was really four times as fast as "consumer" machines, even if it cost fifty times as much, we would buy it. Unfortunately, there isn't. When the product release cycles are favorable, Alpha systems may be twice as good as x86 systems, but not much beyond that unless you are doing the 1k by 1k matrix type stuff.
It is often forgotten that the original Cray-1 was largely a success because it was the fastest SCALAR processor of the time. The vectorization was just a bonus. Now, vectorization is the only thing that gives them a reason for existance.
The tera architecture is very interesting, but for scalar code, it is VERY, VERY slow. I'm not sure if it will be competative with large processor count SGI origin-2000 systems even after it matures. It gains ease of programming from the lack of caches, but it gives away a lot of problem domains where it is going to look stupidly slow.
If a supercomputer company could make a scalar processor that ran many times faster than existing processors and had similar SMP capabilities, it would probably be a success. Even if it cost a million dollars, filled a room, and burned hundreds of kwatts.
The problem isn't really that supercomputers are bad, its just that we are so spoiled by how AMAZINGLY GOOD our cheap consumer hardware is.
I do still worry about the stiffling of innovation that comes from having so few architectural directions for systems, but in the end, wall clock performance is what really matters.
John Carmack
SGI's X server kicks XFree's ass up and down the street. Not to disparage XFree: this is partly because the SGI X server itself is very, very good, but largely because SGI's graphics hardware is better than anything that you can buy in the PC world for love or money. I'm not talking about raw polygons/second in a full-screen window, I'm talking about use as a graphics-intensive desktop. There's nothing that compares to an O2, or even an Indy, as far as running an X desktop goes.