IGNORE FOR NOW: Avid 7-Slot PCI Expansion Unit: N/A

[mapesdhs]

Avid 7-Slot PCI Expansion Unit: N/A (please ignore!)

UPDATE: Upon testing, I kept getting PCI parity errors (with just an Adaptec UW fitted), no idea why, so for the moment please regard this item as not available until I can figure out what's going on. I didn't want to delete the thread though as it ended up being a discussion about stuff aswell.

(NB: in the pictures, the visible option cards are for illustrative purposes only, they are not included by default)

Made by Magma, I obtained this long ago to form a souped up O2 with extra SCSI cards for storage, Phobos FC, etc., intended for some video editing, but was never able to get to it. Perhaps someone else can gain from it instead. It did work ok, all the extras showed in hinv. Certainly ideal for O2 but also other SGIs or indeed non-SGI systems.

There's no key for the front panel, but it doesn't need one, the switch is easily accessible.

Inside there is space for really long PCI cards, good for retro PC builds perhaps.

The pics below show what it looks like with some cards fitted just as an example, then I removed the cards and installed blanking plates in all the slots except one.

Includes the host PCI card. The link cable from the cabinet is 23" long (approx. 58cm). The main cabinet is approx. 17" wide, 17" deep and 7" high (43x43x18cm).

The Adaptec 2940UW cards would be 5 UKP each if desired.

Ian.

And for reference, here's a pic of the four 2940UW cards:

 

[jenna64bit]

Oh that's fascinating, what's the bandwidth like going through a single PCI slot? Guessing that's why you stuck with the 2940UW? I would be interested in this, as currently my O2's disk bandwidth is limited by having the RAD audio card.

 

[mapesdhs]

It wasn't so much sticking with the 2940UW because of bandwidth issues, more that I'd had ideas about how to organise external devices to minimise daisy chaining and other issues. I actually a the dual-channel version of the 2940UW now, ie. it has two external VHDCI ports.

I doubt it would affect the available bandwidth much, it would still be 40MB/sec max per UW channel, but note that I don't think O2 can get more than about 85MB/sec through the whole PCI bus anyway. Thus, best not to think of it as adding to the bandwidth(I don't know if it would help with your audio card situation). The O2 arch diagram claims 267MB/sec for the PCI bus; maybe some DMA ops might be able to exploit a better proportion of that, but anything which has to go to the main CPU will be much slower, especially for an R10K/R12K system since the L2 for those systems is not connected to the SysAd bus. I think the best STREAM result I got for O2 was about 95MB/sec, using a minimum screen res and 16bit colour depth. CRM-to-RAM bandwidth in O2 is fast (as the PR loved to boast), but CPU-to-RAM bandwidth is not.

Actually what I really wanted the unit for was the ability to have the various extra storage devices and suchlike, but also the fibre GigE, but I just never delved further into how the GigE link would work. I did buy the big Netgear switch (listed separately) and obtained the four GBICs for it, but that's as far as I got, as by then I'd been lured away by the faster main CPU speeds of Octane2 and Fuel, which is ironic as I'm moving back to an O2 now for my main SGI desktop (because super quick CPUs didn't really help my main bottleneck, which was javascript performance, something that sucks regardless of the hw spec), though I've not yet decided whether to stick with the dual-channel Adaptec or put the fibre GigE back in and look for a more sensible hookup solution. I did try an RJ45 GigE once before, but couldn't get more than about 35MB/sec through it (I suppose that's still 3.5x faster than 100Mbit, but not as high as I'd been hoping); atm I don't know if the fibre Phobos can do any better.

Ian.

 

[jenna64bit]

Interesting, yeah, in terms of system bandwidth the O2 really feels hamstrung next to an Octane.

 

[mapesdhs]

Sure, but then, an Octane can't easily play back MJPEG, do video as texture, process DV material or various other things without some expensive and rare options. Octane is also loud, heavy and uses a lot more power. They are very different systems, aimed at different markets. To me they compliment each other perfectly.

I recently did some DV processing analysis comparing an O2 vs. other SGIs (full article coming soon). For playback, even a max spec Tezro was still only half the speed of O2, while for export it needed a dual-700 to match a mere R5K/300 O2, with a quad-1gig Tezro just about matching a good O2. On the other hand of course, systems like Octane, Fuel and Tezro were much faster at initially importing DV files, but in real world use playback and export dominate typical usage.

As with so many things, it all depends on the task. O2 is a great all round machine. It is also far more other-half compatible. Small, low noise and "cute" (quoting comments I've heard, lol).

Ian.

 

[telackey]

And the ultimate O2 experience can be yours! Shameless plug goes here: https://forums.sgi.sh/index.php?threads/lnib-o2-r12000-400mhz-1gb-mb-030-1327-002-rev-d.1149/

Seriously though, to me at least the O2 is the best hobbyist / nostalgia SGI. Other than video, I wouldn't exactly say they are stellar at individual tasks, but the combination of all their good qualities is stellar. Especially when combined with the 1600SW display, they are a real winner. I've got 3: an R5K, an R10K, and the R12K O2+ I've got for sale.

Though not the fastest, I have a soft spot for the R5K/R7K series, mostly because the O2 system architecture was designed around that chip. The R10K/R12K were brute force answers that while undeniably faster, were never quite as elegant.

 

[mapesdhs]

I agree, though I do think they are definitely stellar for video, eg. it needs a quad-1gig Tezro to match the DV export performance of a top-CPU O2 of either variant.

I talked to numerous SGI people about the R10K aspect of things. It wasn't something they intended to do, but because it was so much faster, despite the limitations which hindered fp performance, a lot of big name clients wanted them, which was a problem because at first the cache interface ASIC wasn't ready, SGI had to use an expensive standin. Hence why, when IIRC the R10K/195 option was launched (with the proper converter chip), the price dropped by quite a lot. The R10K/R12K side certainly gave it an edge for int ops, and it did bring basic 3D up to par, but the cost was always an issue, and as you say there are aspects of the arch which knit together better with R5K/R7K units, though it's rare that anyone would notice anything relevant in reality. Still, the appeal of the lower heat, no fan noise and extra disk bay are hard to ignore. At some point I want to again look into whether making R7K/600s is viable, but probably not this year, too much else going on, family stuff. I did it once before long ago but it didn't work out well, just too expensive back then.

Ages ago when I was talking to Joe Page about this, he'd asked the likes of IBM about how far the O2 could be pushed. They came to the conclusion that, if the PROM source code was available, then it should be possible to support the dual-core R9K 1GHz or the single-core Sandcraft 1.5GHz, but despite even some SGI employees pushing for the source to be released, in the end management said no. Such a shame. Even ILM was interested in a souped-up CPU module with a big L2 (IBM said 16MB at 50% faster speed would work); quoting the ILM guy I knew, "Compositing in L2!!", or words to that effect. Of course eventually one would run into other limits, the CRM bandwidth, latency aspects, inherant fill rate limits, etc., but still, I'm pretty sure it's possible for O2 to outperform MXI with a good enough CPU, as the GE performance keeps increasing, especially with the R9K, and it would be amazing to have a system that can cope with a heavy geometry load while handling any equivalent fp application load at the same time. It is of course kinda nuts that SGI itself didn't release an R7K/600 and just ditch the R12K, but I think they decided to use the R7K/350 recall as an excuse to stop working on the system entirely, so despite the O2+ launch, they'd already called it a day. I had expected O2+ to use Cobalt or something (from the 320), which has a much higher fill rate. Ah well, the road not travelled eh?

These days all the mod stuff, FPGAs, R-Pi and so on focuses on retro hw from the 80s through to maybe the late 90s, home micros, consoles, DOS/Win95, but maybe in another decade tech will have moved on to the point where it's much easier to reverse engineer the O2 PROM, figure out how to change it to support better options. Alas that level of meddling is way over my head.

Ian.

 

[jenna64bit]

In any case, a PCI box like that solves my 'problem' with an O2 setup. Having to choose either RAD & AV2 or extra scsi bandwidth. I don't know how stressed this would make the PCI bus, but it's interesting. How awful would shipping to the US be anyway?

 

[mapesdhs]

(note the volumetric weight is higher than the real weight; the box would be 55x55x25cm, 11kg, ie. approx. 22"x22"x10", 24lbs)

Shipping via UPS Express Saver (booked via parcel2go.com) would be 75 UKP. I knocked the price down a bit to counter this.

Ian.

 

[mapesdhs]

PS. I was going to list the unit on FB Marketplace and Gumtree on Monday, so if you'd like to get the unit then please let me know by email (mapesdhs@yahoo.com).