Slashdot Mirror
SGI Sets Sights On Turnaround
grub wrote to us about an article about SGI, and its ongoing battles to turn its corporate fortunes around. The company's been doing interesting stuff for a long time - here's to hoping they stay around.
← Back to Stories (view on slashdot.org)
What is SGI's relationship to linux, and why doesn't this article mention linux once?
I'm posting this anonymously because I don't wanna get in trouble in case any of this is confidential. I don't think it is, but you never know.
SGI's official relationship to Linux is this: none whatsoever. At one time, there were some pretty serious plans to port core OS libraries, build abstraction layers and shims, and phase out the IRIX kernel in favor of Linux. It was thought to be easier to turn Linux into a real commercial-quality kernel (not my words; don't flame me!) than to port 60-odd million lines of IRIX to the then-proposed IA-64-based Origin 3000 variant. These plans have been informally shelved, meaning SN-IA is still a product on the roadmap, but nobody is working on it. It seems to have been put in the "maybe after McKinley" category.
(Take all of this as unofficial comment, of course, but this paragraph is total speculation on my part. I wonder if part of the reason IA-64 isn't really going anywhere in this market is because of the lack of a really good Fortran compiler for it. The MIPSpro Fortran 77 compiler, which I've worked with a lot, kicks ass, and I understand the F90 one does as well. Getting all of that tuned, optimized Fortran code to run on IA-64s seems to be a challenge for a lot of folks that are long-time die-hard SGI customers.)
SGI is officially committed to continuing to develop and ship systems based on the MIPS processors (R12000, R14000, and the upcoming R16K, and R20K) with the IRIX OS until further notice, which is to say that they're not opposed to exploring other options, but there just isn't any reason to switch that plan right now. The Origin 3000 server is a great piece of work, and the new lower-end Origin 300 is selling nicely, too. On the workstation side, believe it or not the humble little O2 is still selling briskly-- now with R7000 or R12000 processors, painted purple, and called O2+. Octane2 kicks ass, and a new workstation to be announced in January or February is also going to be based on MIPS/IRIX, combining Octane2 graphics with Origin 3000-style architecture.
So the official story is MIPS/IRIX forever.
Unofficially, SGI loves Linux. Check out oss.sgi.com sometime to see what SGI is doing with respect to Linux specifically (XFS, kdb, bigmem, NUMA, STP, etc) and Open Source in general (GLX, Inventor, Performer, etc).
SGI is caught in the classic problem that killed DEC, and is killing Tandem, Stratus, DG, and many others: the performance of the lowend is improving so quickly that we can do things on $1K machines that used to require $1M machines.
I have a friend who had an idea that could have saved them. When he was at SGI, he pointed out that machines that were optimized for graphics had to have great I/O performance, which would also make them great performers in another I/O intensive task: running RDMS engines like Oracle and Sybase. SGI management wasn't interested.
So, SGI employees and stockholders lose out, and the rest of us gain another lesson in the dangers of rigid thinking.
-jcr
The only title of honor that a tyrant can grant is "Enemy of the State."
SGI used to be a real innovator in the field of graphics. Now it seems like companies like ATI and Nvidia are actaully doing more for that field.
You know, that's the funny thing about SGI's graphics hardware. InfiniteReality graphics first came out in January, 1996. Since then, SGI has put the same graphics processors on a new system interface for the Onyx2, and tweaked some components in the system twice (called IR2 and IR3, even though the changes were very minor).
InfiniteReality--apart from having the coolest name of any graphics subsystem--has remained essentially unchanged since 1996. IR today has slightly faster geometry processors and much more TRAM than the original IR, but in every other way it is identical.
That's six year old technology, baby. And the rest of the world is just now starting to catch up.
Guess that's why SGI has been selling the same graphics hardware for all this time. Because they can.
But the thing that sealed their doom was when they didn't take the opportunity offered by Nintendo purchasing a huge number of R4000 chips.
The Nintendo 64 was built around the MIPS R4300i embedded processor. It had limited R4000 instruction-set compatibility and a 64-bit execution unit, but it wasn't really related to the chips SGI used at the time at all.
So while selling tons more embedded processors would have been a nice bonus for SGI's MIPS subsidiary, it wouldn't have affected their core business on bit.
SGI has always had a hard time trying to market itself. They've typically made endless incorrect assumptions and end up preaching to the choir. And yet, the wow factor that made the company and it's little cube logo a legend in the late 1980s is still there, abeit in a slightly different manner.
True, not everyone needs 512 or 1024 CPUs running on a single system under a single kernel. Or 16 graphics pipelines. But there are those that do. Which is why, shortly after the introduction of the Origin 3000 two years ago, an entire convoy of the machines were sent to Fort Meade.
It's almost as though SGI has gotten used to the high end, as though their technology (HW, SW, APIs, SDKs) no longer impress themsleves. Nowhere else, not even E&S, can a person find a platform that can drive up to 128 display channels (16 pipes x 8 channels per pipe) with perfect sync, or even at all. O2K and O3K (and more recently, O300 and Octane2) can drive multiple displays from one or more graphics pipelines. Raw, per-CPU performance isn't anything to write home about, but the thruput and latencies are perfect for generating insane 3D and mixing it with streams of HDTV... or anything. Think of a way-cool use of video and 3D. Now increase the complexity and choose, oh, 4 camera viewpoints. Maybe an additional display for stats and another for an "operators station". Easy with O2K/O3K (aka "Onyx" when gfx are invloved). It can be done and it's proven. They've been doing this sort of thing since you and I were using our "cutting edge" unaccelerated 2D graphics cards running at an "insane" 1024x768.
A pair of old demos SGI likes to show off are sometimes called "from space to your face", in which over 500 GB of sat photo textures are shuffled thru one or more InfiniteReality graphics pipes to provide a realtime "bungie jump" from the moon to earth and back. INSANE. 60fps/60hz locked. 4 huge disk RAIDs composed of dozens of drives grinding away like mad to keep the textures coming. WILD STUFF. All in a day's work.
SGI isn't about buzzwords or about wizbang marketing. It's about providing modular solutions to some of the most challenging problems. They've been there to provide HW and SW to those wishing to work on the cutting edge. In 1988 they were selling 3D workstations. In 1991 folks were doing crazy 3D and video mixing. Today their hardware can be used to drive gobs of displays and to shuffle huge amounts of data. Sure, the desktop PC in 2007 will be able to do the same thing. By then, PCs will be able to drive gobs of high end gfx subsystems, and even a cheap graphics card won't sneeze at several GB of textures loading and unloading every second... but until then, for those that need this TODAY, there's SiliconGraphics.
Let's hope SGI is here tomorrow to show us even more cool things.
>Cutting edge graphics, where did you go? Please tell me there's more to the 3D world than IR,
>WildCat II, and GeForce3. Has *nothing* (other than cost) really changed over the past five
>years? It's almost as though I haven't missed anything in the 28 months I've been away from 3D
Dependent texture reads are the only really new thing in the last year or two (and only really got worked out right in the Radeon 8500), but next year is going to see floating point pixel formats, which was going to be one of Bali's truly important points. We should also see highly scalable boards built on consumer chips, which has been promised for years, but (with the exception of some 3dfx high end systems) not delivered properly.
John Carmack