SGI: Hardware

It's old news that in the visual effects industry, MIPS/IRIX is effectively dead on the desktop. There may be post houses with legacy machines still in their pipeline, but what about new hardware? I'm curious to know: Aside from discreet, who is buying MIPS/IRIX workstations in 2005, and what applications do they use? Obviously SGI still has a big presence in high end visualization, HPC and defense, but that's primarily driven by solutions on the Onyx/Origin level, not by workstations. So who's buying/using Tezro (besides discreet), and Fuel?

Let's extend the question further: In 2005, who buys UNIX workstations in general (IBM, HP, Sun, etc.), and what applications do they use? How are UNIX workstations able to compete in price/performance against the Wintel PC juggernaut in any industry anymore?

I assume that in certain industries, there are still valid reasons for choosing workstations over PCs. I also assume that in some instances, it's simply a case of momentum, that an industry has so much invested in legacy UNIX architecture that it's easier to replace existing hardware with the same thing than to convert their entire infrastructure (hardware and software) to PC-based solutions. I'd like to hear from people working in industries where UNIX workstations still endure: what are some of the reasons PCs haven't taken over yet?

Finally, there is speculation that SGI is working on a new Linux workstation, below Prism in the product line. Who would buy it? How could it justify its price/performance over commodity Linux PCs? What Linux applications could require greater performance than that attainable on PCs but less than that attainable on Prism? Did SGI fail to learn their lesson about the commodity market after their disastrous experiment with PCs?

Anyone?

-zolo

In my world:

- CAD/CAE desktop workstations (I-DEAS, UG, Pro/E, Nastran, ANSYS, ...)
- CAE compute servers (Nastran, ANSYS, ...)
- Large scale distributed applications (Oracle, PDM servers).

In this world, PCs usually cannot compete. They perform better for small problems, but cannot scale up past 32-bit addressing...

As for the Linux workstation, I don't see it being much used in my world as most commercial applications I use are not yet supported on Linux and probably won't be for a few years...

a sgi linux desktop is probably a must. with the end of life for mips based systems and their migration towards linux they have to offer something on the low end.
discreet seems busy transferring their applications to the linux platform as well but with pc hardware they still face the bandwidth problem as it seems..

as long as an sgi offering offers substantial advantages over a standard pc design, that cannot easily be replicated there will always be a market, even at their pricing structure. they should have the knowledge to build a media workstation that's geared better towards data intensive applications than a beefed up off-the-shelf-pc (read: intellistation) and that's where they might be able to attract new customers and maybe gain some marketshare back from where the intellistations and boxx's are nowadays.

as for their previous pc offering: i'm not aware that the visual workstations were anything else than a somewhat modified pc. they definitely didn't stand out enough from the competition to justify their price since there were no significant unique features (imho). still, they sold but that was probably due to the (then) famous brand.

ANSYS and Nastran are fully supported for Linux and workstations running Linux, both 32 and 64-bit.

Yes they are... 32-bit Linux needs a recompile of the kernel to address more than 800MB, but it works a little better than on Windows...

"Fully supported" is a big overstatement if you're used to dealing with software companies. MSC supports 2.4.x kernel, UGS supports a little more, including sgi ProPack. This is certified OS. It doesn't mean it won't work on more recent versions, but it means the software vendor is not obligated to support you if it doesn't work... ANSYS supports about the same as UGS as far as Linux OS.

A good example of sgi Linux technology: NX Nastran is actually optimized for the Altix platform, based on sgi specific libraries ported from IRIX (ffio, MPI, ...).

What's not running at all or not very well on Linux are the major pre/post processors... Patran runs like shite on Linux, FEMAP and I-DEAS don't run at all. These three account for about 85%-90% of the FEA pre/post software... One of the biggest issues with Linux is getting consitent OGL performance and quality. These tools heavily depend on OGL.

A big difference between the Intellistation/Boxx and a new SGI Linux workstation is that the former both run Windows, which account for the vast majority of their user base. Until more professional media applications are ported to Linux, a new SGI Linux workstation won't have anything to run on it in that market aside from discreet apps and Maya.

The Visual Workstations, at least the 320/540, were a case of too little, too late. While their IVC architecture was logical, elegant and indeed admirable on paper, by the time they were released, their performance didn't justify their price premium. Had they been released two years earlier, they might have seemed more impressive. However, their hardwired graphics, proprietary memory and reliance on hacks to the existing Windows NT code would still have been an issue for many customers. SGI didn't grasp the fundamental mentality of the commodity PC user base--cheap, standard user-upgradable hardware is what drives that market. It's a really tough market in which to innovate, unless your company is named Intel.

In the nuclear world Unix workstation are still used for a couple of reasons:

1.) All programs were originally written for Unix, and the QA work involved
in verifying a compile on a new platform can be a great hassle.
2.) The documentation required for QA is enough already without having an
operating system that is still evolving, once a system is together it stays
that way.
3.) The hardware prices do not seem to be an issue for big shops that
just need it to work, plus no one feels the cheaper stuff is robust enough.

Also I have seen many cases in CAD where, yes it is supported on x86 platforms,
but they simply cannot keep up with the RISC sytems.

PC have there place, but it is not everywhere (thank goodness).

hope this provides some insight!

---
Charles

Maybe you don't count them as UNIX but Apple Mac OSX is unix with a sweet GUI.
They have some pretty nice products on the desktop and their xserve and xserv Raid look like pretty nice products in the server market. Their new SAN file system looks pretty nice too for the film and TV post business. Time will tell.

As for IBM, HP, Sun etc. I don't know maybe we should say "count the days." Their market seems to be going going gone. Even though CAD is a bigger market than 3D Fx I don't think its a big enough market to support new hardware development by itself.

Are you in the nuclear world Charles? Alway's nice to meet another person that glows in the dark Here's me...

Yes I am in the nuclear world! You wouldn't happen to work
for Susquehanna would you? (it's one of the plants we work
with)

I am not so lucky as to get that close though, most of my
work is in analysis.

---
Charles

Been there once. I've been hitting the re-fuel shutdowns at Limerick since 96, that is where that picture was taken. I should be heading there within the next few days for pre-outtage work. I've also got two re-fuels at TMI under my belt, and one at Beaver Valley, but don't like to do the traveling thing much anymore, luckily Limerick is only 40 miles away

We are actually doing some work with Limerick right now as well.

So as not to get too off topic, check your inbox.

---
Charles

Sun and HP, who knows ? (Itanic seems to be tanking tho ) but the PowerPC doesn't look like it's going away any time soon ... All this talk about "markets" and other salesman-speak - Gack ! Look at the figures unixmuseum posted a while back. For FEA work on largish models the whizz-bag peecee was four minutes faster than an antique Octane. Wow. There is so much hogwash in the computing world. I don't think you'll see the people with brains changing to peecees until the companies supplying Unix workstations quit making them due to believing other people's marketing hogwash more than their own marketing hogwash. (Although I've been known to overestimate the intelligence of the American public before.) Marketing is going to have a lot to answer for in the next life.

Then we'll have no choice - but until then no one with a brain wants to run a Windows operating system in a professional capacity.

Let's not even discuss the crappy quality of commodity peecee hardware and/or the never-ending driver upgrade treadmill or the fact that you can't find the drivers for two-year old hardware half the time or the way that the hardware changes without warning so that driver A doesn't work with hardware A unless you have rev 16.39678B-12971 but that's been removed from the website (or was never there) because we "don't support legacy hardware due to cost considerations altho we were plenty happy to take your hard-earned money just last week for this card which we touted as the best thing since sliced bread and the Operating System for the Nineties but that was when we believed we could sucker you all in - since that didn't work, tough luck" etc etc ad infinitem. And the OpenGL issues ! Look in the Pro/E newsgroups. At least once a week someone asks a question which gets the response "you need a professional-grade graphics card for $2,000." So where's the fricking cost advantage to peecees ? You've just stuck a $2,000 agp2 card into a $50 piece of crap ? An agp2 card that's gonna be "legacied" into oblivion by agp4 in three weeks ? Did y'all know that perfectly good AGP cards won't fit into newer motherboards already ? WTF ? this happened within just a couple years of the introduction of AGP ? Who wants to spend all their time buying and maintaining hardware ? And how cheap is it if you have to spend two months per year fiddling with it and replacing it? And who the hell wants to depend on what Microsoft has up its sleeve for the next Windows release ? "All your data are belong to us." Oh goody . The bad-drives-out-good syndrome is at work within the peecee world just as it is in the peecee-workstation conflict.

If there's no place in the future for quality Unix workstations we'll all be a lot poorer.

Here are a little more results, as I recently upgraded the main Octane2 to a dual 400 (which leaves an extra single 400, if anybody is interested).

The model I posted results for wasn't a big model, but yet the numbers kinda spoke for themselves already:

- PIV 2.7GHz: 76min to complete the task
- Octane2 single 400: 80min
- Octane2 dual 400: 65min
- Tezro dual 700: 44min

I should go through the effort of doubling or tripling the model size, the gap would be even bigger (especially when we hit that 32-bit limit )

To me, the bottom line is not simply the time it takes to accomplish one single task. It's "how many tasks can I accomplish at the same time?".

Well, with my groovy PIV, I could accomplish a grand total of 1 task during 76 minutes: solving the model. Even the screen saver (nothing eleborate, just the WinXP logo in 2D) was putting the whole system in jeopardy!

With any of the UNIX workstations (especially the dual CPU), I was working on other things at the same time... This example is with sgi, but it would be rtue with HP or Sun...

I still have many of my clients still buying UNIX workstations (and definitely servers) for that exact reason. The first few days they're being picked on by the PC boys who have PIV 128.234THz (the 128.235 is coming out next week, let's replace the old 128.234!) running small test models soooo much faster than the UNIX junk... A week later, they come by the UNIX guy, begging for CPU time as their PCs can't take the large models...

I'm not saying there is no place for PCs, I'm saying the PCs do not and cannot replace everything.

I set a can down on my Octane dual-300 keyboard 'enter' key the other night by accident and inadvertantly opened about 40-50 copies of RealPlayer, which all chimed perfectly and loudly without so much as a hiccup - the thought that went through my head was "Try *that* on a PC"

We are preapring to buy 5 new Tezros as hardware-in-the-loop because of (in order):

a) non-preemptive multi-processing (i.e. REACT real-time)
b) scabable symmetric multiprocessing
c) stability
d) 64-bit extended precision

are you aiming at the single user/small studios here? i agree, as long as some standard apps do not show up on linux, there's no point in going 100% linux for these people. they would need a sidekick windows or mac machine for running certain tools which is hardly acceptable.

however, intellistations and their likes are nowadays running linux in larger studios. a lot of 3d cg related stuff has already been ported, not just maya - which is the standard, btw so for big vfx studios which have made the transitition from irix -> linux, i'm sure they would be interested in sgi offerings again.

I don't know if moderator maby works at NASA (? I'm guessing) so then has different ("special") relationship with sgi and support, but let me tell you what happened when Fuel and Tezro came out and I called local sgi representative to ask about the price..
I called,.. secretary answered, I asked about Fuel and Tezro price and then she said to wait a moment to give me the guy who is in charge of sales on the phone - then he answered and when I asked about the new Fuel and Tezro workstations his words were, if I translate them to english: "..oh, no, no,.. it's really pointless to buy that 'cause new P4 will perform much better than those workstations. The new P4s, AMDs and those new graphics boards that you can buy everywhere can already do all that those workstations can.. Thank you for calling."

That was FU****G OFFICIAL SILICON GRAPHICS COMPUTER SYSTEMS SALES REPRESENTATIVE!

This is a pure "school example" story that I tell everyone when asked about who buys UNIX workstations these days. The fact is that major manufacturers - HP, SGI, SUN, IBM, NEC,.. have already decieded that new future direction is going to be Linux, Windows, Solaris, AMD, Intel, POWER, PowerPC, Ultrasparc and if you show any deviation from that direction then they will, over time, FORCE you to buy it. There are no more secrets about that, these companies have "quiet" agreements between themselfs about market control.

Slight tangent here, but if you look at UnixM time results:

- PIV 2.7GHz: 76min to complete the task
- Octane2 single 400: 80min
- Octane2 dual 400: 65min
- Tezro dual 700: 44min

You will notice that the Tezro Dual 700 timing is pretty much exactly what you would expect if you just had single CPU ie 80 * 400 / 700 = 45 min

The difference between single 400 and dual 400 likely represents moving of OS processing (IRIX) to second CPU making other one available just for model processing.
If UnixM model process application was multitreaded then I would expect that time would go way down: 80 * 400 / (2 * 700) = 23 minutes.
So to get real performance improvements you need to look at the application software not the hardware.
Also just want to say that when it comes to interactive response multiprocessor IRIX machines are much much better than Windows machines. In fact I think it is safe to say that Windows NT based OSes are so inefficient that most of the processer performance improvements make by Intel get consumed by the OS.
I think that it is in the weight of the OS (ie how much cpu power is consumed by OS rather than available to application) that Unix (IRIX, FreeBSD) still has huge performance advantage over Windows.

It is true that in general, at least my experience, that the computing power on the sgi increases linearly with the processor speed. This is quite an accomplishment, something we don't see much anymore...

But in the case of this Nastran solve, it's a little bit more complicated than this. A Nastran solve involves a little more than pure CPU, there is actually a large amount of I/O going on.
Nastran is "multi-threaded", in two ways: SMP & DMP. SMP, which is the way I ran this model, doesn't bring that much to the table, only a few percents. DMP, which is a cluster-type distribution would definitely bring the compute time down. It's not a simple linear performance improvement, it never is, although it is close with DMP. DMP can actually be used on a single machine, multi-CPU machine, and will bring better results than SMP. It also greatly depends on the types of problems being solved

Other benefits that I see in the sgi architecture (and these made it into sgi Linux Propack): dynamic threads, to be able to suspend a process and come back later and start it again. This functionnality is quite amazing: you've got a process that runs for 3 days, you stop it and perform system maintenance and whatnot, reboot the machine, do whatever you want. You are then able to restart the process where you left it! Another one is ffio, an asynchronous read/write of scratch files... This by itself usually gets me 15%-20% time gains. These are unique to the sgi architecture...