robthebloke writes:
> nah. What put the boot into SGI systems was their premature jump to
> Intel Itanium processors. ...
Alas, not SGI's fault. Itanium was very late, and the first iteration
was not that good. This is why the original Origin2000 rack could
only hold 16 MIPS CPUs, because it was designed with the thermal
characteristics of Itanium in mind (much hotter). Years later SGI
released Origin3900 with up to 128 1GHz MIPS CPU in one rack, showing
just what was possible given the lower cooling requirements. But by
that time it was far too late of course.
Sadly this was a transition they could not prevent. Many of the SGI
CPU design team left in the winter of 1996/7 (for Intel or
elsewhere), drawn away by huge salaries. Intel played a skillful game
of bragging about how good Itanium was going to be and thus obtaining
but by bit the very design staff they needed (from SGI, DEC/Alpha,
etc.) to design Itanium. I knew someone on the R10K design team back
then; he said even floor sweepers in the CPU labs were paid enormous
amounts. As a design though, the current Itanium2 is quite good; John
Mashey (STREAM author) told me that many aspects of what SGI had been
planning for SN1/2 (Beast/Alien) are incorporated into Itanium2, eg.
the way it performs so well despite its low clock, though the chip
doesn't have the additional vector/media extensions which SGI was
working on (MDMX, etc.) SGI wanted a chip that offered vector
processing in a manner that allowed each register in a vector to also
be used for media processing - this would have given enormous
processing speed for video, etc. But without the staff, moving on with
these ideas was going to be difficult at best.
I was told that if SGI had continued with its own plans, and at the
time assuming Intel was able to release Itanium as planned, SGI's CPU
would still have been faster (by about a third or more) but too
expensive to compete. Thus, at the time, the switch to Itanium. I very
much doubt SGI would have switched if they'd known Itanium was going
to be so late. For a start, Origin would have had a much greater CPU
density, greatly increasing its price/performance ratio.
That wasn't really what ruined their business though. I've dealt with
a wide range of different companies over the years, helped sell $M
worth of SGI kit. IMO the 3rd-party sales model was very damaging,
tech info was not available easily enough about key aspects of
certain products, the marketing/PR was often woeful (which led to
poor targeting of customers for products, eg. selling O2 as an
upgrade over Indigo2 which it was never meant to be - this happened
at ILM), and numerous times, despite feedback/comment, they simply
didn't produce what customers wanted, eg. a speed-gfx solution
without advanced features such as 48bit (aimed at CAD/animation)
instead of the high-quality gfx designs like VPro. Likewise, lots of
daft design choices like Fuel not having audio or Gbit as standard,
Octane2's mbd not having U160 or Gbit, the later PCI cage for Octane
not using 66MHz PCI and a better XIO/PCI conversion ASIC, not using
the Cobalt core for O2+, not changing O2's PCI to 66MHz in the O2+,
not changing the Ethernet to Gbit or the SCSI to U160 in the O2, not
filling the other 4 GE sockets in the IRx GE board to double its GE
speed, etc. I could write a huge list, but you get the idea.
Point is, it was a combination of things. From secretive pricing to
bad marketing. They had excellent products which were poorly sold.
Once the rot set in, key staff left, including most of the two dozen
or so people I got to know over the years. NVIDIA did its best to
poach the gfx division, which worked very well for them, the original
GF256 being based on IR gfx. Irony is, ATI nabbed some of the same
people from NVIDIA later on. But even then SGI had opportunities to
get back in the game, eg. NVIDIA wrote drivers for one of its Quadro
cards for Octane, but SGI said no to releasing it. Big mistake.
Btw, the current VP Product Marketing at NVIDIA (Ujesh Desai) is the
person I reported to in 1998 when I had an O2 loaner system from SGI
and was doing IRIX 6.3/6.5 testing after launch (bug fixing, especially
mainly on the media tools). He was O2 desktop manager at the time.
Funny how things work out. :D I still have the O2, customised and
maxed out now in various ways of course.
> hastily crapped out a new mips CPU on their Fuel [sgi.com]
> workstations. We didn't buy them, because we were waiting for the
> Itanium ones. ...
Fuel is a perfect example of the bad marketing but also of design
mistakes. It was not meant to be a new low-end, but some perceived it
as such. Likewise, others regared it as an entry mid-range, which it
largely was (aimed at companies who did not need more than one CPU)
but it lacked certain key things, eg. a speed-gfx option. For HPC and
other markets, the RAM limit was too low - one company told me it
would have been ideal for ANSYS, but the max RAM needed to be more
like 16GB or 32GB, not a whimpy 4GB. Likewise, in theory it can be a
nice entry Flame system, but only with V12 (which lacks raw speed;
V10 doesn't have enough VRAM) and then again the system RAM isn't
that good for working with HD anyway. Don't get me wrong though, I
love the Fuel system - I use one as my main daily-tasks machine
(typing on it now; R16K/900MHz 8MB L2, 4GB RAM, V12, 300GB 15K, 300GB
10K, 1TB SATA, etc.) - but it could have been sooo much better. Not
coming supplied with Gbit Ethernet and proper audio was silly. It
meant adding them used up PCI slots which one would normally prefer to
use for something else. At least Tezro came with audio and Gbit
included (I have two, including a quad-1GHz).
> ... So they switched to Intel Xeon [nitroware.net] CPU's
> running NT, and we didn't buy them, because as we know, the Itanium
> hit problems, and a dell workstation running linux was a cheaper
> option. ...
Irony is, the first VW system with XEONs (540) was very good and much
liked in industry, especially in the MoD where it was used for
aircraft design. Likewise, the 320 was very strong for similar
reasons - it was the first 'PC' I ever saw that could load a 50MB 2D
image in less than a second, and a dozen of them were used in Bosnia
for coping with the UK force' satellite image analysis requirements.
But again, not enough tech specs were public, designs mistakes were
made (using proprietory RAM was dumb) and SGI didn't follow through
with their original roadmap for evolving the IVC architecture. The
MoD and other corps were expecting the next system to be an 8-way
machine with the same architecture, more RAM, faster gfx, offering
the same advantages as the 540. Ditto a 320 followon. Instead, SGI
switched to conventional PC build for the 230/330/530 systems which
offered nothing unique at all, and hence no way to compete in a busy
market and not what customers wanted. At the same time though,
resellers were selling 320s as render nodes which was very silly,
raw CPU/RAM speed not being a strong point of UMA designs.
I once ran a design lab of a dozen 320s (2000 to 2003), mostly single
PIII/500 with 512MB RAM. For a large urban model with a lot of
texture (almost 200MB) they gave about 10fps. When upgrades began to
become necessary, there was no proper next-gen system on the horizon
and CPU/RAM upgrades were costly (reseller markup issues again), so
the lab switched to P4/2.4GHz PCs with GF4 Ti4600s. On paper, the P4
PCs should have been much faster. In reality, they were disastrously
slower at first, not even 1 frame per minute, because they couldn't
handle the massive texture thrashing that was occuring due to the large
texture data sets involved and the widespread use of 16K composite
textures - issues which make no difference whatsoever on the 320/540.
The lab guys had to redesign their models with the limitations of PC
gfx in mind: no more composite textures, lots of LOD built in, etc.
They sorted it out eventually, but just a shame SGI didn't stick to
their original plans and release a 2nd-gen IVC line.
By way of an example, O2 can in theory rotate/pan/zoom a 700MB 2D
image without falling flat on its face. I tested this once and it
worked fine (16000x16000, 732MB). Pretty impressive for a system with
only a 300MHz CPU - try it on a normal PIII/500 PC and see what
happens! :D
O2 is a really nice system, ruined eventually by bad marketing and
lack of appropriate upgrades.
> ... Over the course of a couple of years Sgi machines literally
> vanished from the Cg industry.
It wasn't quite that fast. Heck, I'm still supply parts/upgrades to
companies today. But yes, overall they vanished pretty quick where
there were not so many cost issues with switching to Linux/Windows/Mac.
Where speed is not a factor though, SGIs are still used, eg.
controlling printing systems, textile manufacture, image processing
(ie. industrial process control) A lot of enquiries come from Asia
these days, and even institutions such as the Australian airforce are
still using flight training systems running on Indigo2/Onyx2 (I've
sold them parts several times).
> al). The comsumer grade Geforce cards had better OpenGL support +
> features than an Sgi unit at a fraction of the cost.
The GF cards had the speed, but the image quality was awful back
then, especially the mipmapping. The fact that this didn't really
matter so much (like you say, these issues were sorted out
eventually) was the bandwagon SGI missed jumping on by a mile. In
2001 I tested a dual-400 Octane2 V12 vs. a GF4 Ti4600 PC P4/2.4
(various Inventor tests); main CPU power notwithstanding, the GF4 was
6X faster than the V12, but the mipmap quality and geometric accuracy
on the PC was really bad. The urban lab design people saw similar
issues - wobbling geometry, poorly aligned verticies, etc. But the
cost differences were impossible to ignore, and there was always the
Quadro line if one wanted better quality (still cheaper than an
Octane2/VPro).
It may have been a different story if SGI had made use of one of the
better Quadro boards for Fuel/Tezro/Onyx3K. When they did finally try
using such a GPU (Prism, etc.), they went with ATI which was a
disaster - the drivers were nowhere near reliable enough.
> SGI and Apple. Apple (having seen a computer company crash and burn
> due to a switch to Intel) must have studied what went wrong with Sgi,
> and made damn sure they didn't repeat the same mistakes....
That sounds very plausible.
> ... If Sgi
> had managed the transition as well as Apple, it would still be a
> powerhouse in the industry.
SGI had so many chances to stop making the same old mistakes, but yet
more dumb decisions kept being made, despite all the advice & comment
coming in. I forwarded comments from companies and users all the
time, including from the likes of ILM - one of their staff (a tech
director for Star Wars Ep. II) told me they would have loved to have
seen a proper 2nd-generation O2-type system with the same architecture
but a lot more RAM, faster gfx (eg. Cobalt, which is 10X faster than
O2's CRM), Gbit, U160, PCI64 or PCIX, something like the dual-core
1GHz R9000, etc. They said the O2 was great for compositing and other
things (how we forget that O2 held the records for Shake/Chalice for
a while), but without updates, etc. it floundered compared to Linux
systems. With nothing new to properly replace it, they had to move on.
The SGI hobbyist community has done much to make the most of what
traditional SGIs are capable of, such as the R7000/600MHz CPU mod for
O2, changing Origin350s into Tezros, etc. And it's easy now to obtain
good spec MIPS/IRIX SGIs, especially high-end systems (I have a 36-CPU
Onyx3800 2-pipe IR3, plus a quad-500 Onyx2 IR4 deskside). So even
though such funky systems are not made anymore, one can still indulge
in a spot of nostalgia. :D And in some cases they can still be
perfectly useful systems, Fuel being a good example. I'd never use
my Windows PC to access internet banking... :|
The old SGI is gone. Rackable likely bought SGI for the branding, so
no suprise they'd use the Octane name in this way. At least Octane III
is not some generic 2-CPU XEON box like all the other offerings from
Dell, etc. However, I would like to know why, with gfx included, it's
limited to just 2 CPUs. Seems a bit odd.
Ian.
http://www.sgidepot.co.uk/sgi.html
