SGI: hinv

Fastest SGI MIPS CPU ever, plus a V10! Still looking for a V12... but too broke to buy any

hinv:


and gfxinfo


recondas also asked me if I could post some more info since it's quite a unique beast - so without further ado-

serial all:


env:


In addition, I discovered something new about the Origin 350: It can have one of three types: Chimera Server (Origin350), Chimera Blade (Onyx350 IP?) - happens when V10 is inserted, and Rackmount WS (I assume this is a Tezro) - Activated by inserting the V10 and running "make rmws 1" on the L1.

This status can be viewed from the brick command, for example:


I was wrong - Rackmount WS sytems *can* learn serials from L2s - as long as the serial begins with a P rather than an M. Thanks to recondas for noticing this.

When a headless O350 has make rmws 1 set on it, its type is still "Chimera Server" but it gains the ability to also accept a P-prefix, allowing it to be linked to a headed system for a NUMA-linked Tezro. Sadly I think this configuration won't extend beyond 2 systems due to r-brick serial security - but I may be proven wrong again. I don't have an R-brick or enough O350-bricks to test with.

Details on the V10 hack can be found at: viewtopic.php?f=3&t=16719768 for the curious.

That rocks, good luck with pairing them up and with the hunt for graphics. You should be able to get a G-brick from an Onyx 3000 attached to it for graphics, or a V-brick if you can find one. It would be good to see how the CPU's perform on some of the benchmarks here. I just did Ian's Blender http://www.futuretech.blinkenlights.nl/blender.html benchmark on 8 CPU 500Mhz Origin 3200 (With V-brick) and got it to 61 seconds. I reckon that you might be able to shave a few seconds off that with the additional cache you have (But fewer CPU's). But then again, you've not got graphics yet, so we'll see......

Good luck though!

Very cool!

Sounds like a fast system!

For those that don't know, I managed to mess my O350s up something amazing:

viewtopic.php?f=3&t=16719038

Thankfully I should have an L2 on the way soon and I'll be able to fix everything up (still looking for the L2 emulator, though, because it'd be nice to have around)...

Anyway, attaching a pic...

I hope you can get it working again! Would be interesting to see how it compares to Toby's 8x500, though I don't think the
larger L2 will make much of a difference since the Blender test scene is pretty simple. I've been running a complex Alias test
which benefits from a larger L2 though, eg. Origin300/600MHz beats a dual-400 Octane2 (1/3rd of the speed difference is
due to the faster O3K RAM, 2/3rds of the increase is due to the 2X larger L2). By contrast, a different simple Maya test
scales more with basic clock speed, eg. Octane, Origin and Fuel are all pretty much the same with a 600MHz CPU.

The more complex the task, the better it will run on an O3K system.

One thing that does surprise me however, and a bit of a caveat for your O350: I see from the hinv that the L2 runs at
333MHz. With a 900MHz system, it runs at 450MHz. This might make a difference for some codes. Anyone have a
quad-1GHz Tezro to compare? Strange that SGI couldn't just make the L2 run at 500MHz.

Ian.

PS. Assuming you're successful in your repairs, any chance of some C-Ray results, etc.?

I'll benchmark the crap out of it as soon as I get it working again (whenever the L2 arrives... should be this week or early next...)

The L2 speed discrepancy is curious... I wonder why they did that, maybe the L2 barely doesn't run at 500Mhz or something so they had to use divisor 3 instead of 2 (that's totally stupid if it's true though)...

I might be getting IR4 for it as well (although it would probably end up being IR2 since part of me getting the IR4 involves selling the RM and getting a cheaper one)... we'll see how that goes.

bri3d wrote: The L2 speed discrepancy is curious... I wonder why they did that, maybe the L2 barely doesn't run at 500Mhz or something so they had to use divisor 3 instead of 2 (that's totally stupid if it's true though)...

Indeed. The L2 can run at 450 ok, so not much of a stretch to think of it running at 500. For some codes, the lower L2 speed might
be enough to offset some of the benefit of having a larger L2 and/or higher clock, eg. compared to an 800 with 8MB. Hard to say.
Does anyone have a Fuel, Tezro or Origin/Onyx3K with an 800 CPU? What is the speed of the L2?

For reference, a dual-700 Tezro uses 350MHz L2, while I already know a quad-1GHz Tezro also uses 333MHz L2. Rather ironic that
a Fuel/900 has a 35% faster L2 link.

Ian.

So 800Mhz = 400Mhz L2.

and as Ian says :



So 1Ghz = 333Mhz L2.

Ah well.

Anyone know what kinds of tasks/codes might be affected by the L2 speed? My site doesn't have any fp intensive tests
as yet, though I'm working on it, eg. Autodock, Alias, etc. Even then, fp tasks vary greatly as my SPEC95 pages show.
Perhaps hard to tease out the effects when, despite being slower, the L2 is 2X larger.

Ian.

They're baack!

Got my L2 controller hooked up (after finding the power pinouts) and gave them new serials - running like a charm.

NUMALink still won't work though - I blame my cable; it's a NUMALink cable from Origin 2000, and I think NUMALink 3/4 cables might be different somehow. The L1s see each other and * pwr on from one L1 will power up both (even without the L2 involved or connected), but when the PROM probes NUMALink it just says "Network down" and only finds one node (itself). I've tried clearing partitions and everything but to no avail

Weirdly there's nothing about any kind of NUMALink error in any error log... you'd think that if I had a bad cable it either a) wouldn't work at all or b) would error somewhere, but instead it just works until the PROM tries it, then silently doesn't find the other system.

Here's what the L2 and L1s have to say:


So the L1s and L2 are definitely communicating amongst themselves properly - the L2 sees both L1s, the L1s see each other as attached C-bricks via NUMALink, and all should be well.

But when I issue a pwr up, everything appears to be working until...



And then both PROMs operate independently of each other and the systems don't acknowledge their connection (which confuses the crap out of the L2 system console function as well).

I'm sure Toby can confirm, but yes, I would have thought you'd definitely need to have the O3K-type NumaLink3
cable to link the two O350s together.

Ian.

bri3d wrote: They're baack!

NUMALink still won't work though - I blame my cable; it's a NUMALink cable from Origin 2000, and I think NUMALink 3/4 cables might be different somehow.

Aeh... cant believe how you are able to connect a NUMALink 2 cable from an O2xx(x) to an O3xx based system

Numalink for O3xx(x)

They have a huge massive connector


Numalink for O2xx(x)

The pin layout looks similar but the connector doesnt match to another type.

Good to hear that your machine is back from the dead .

Regards
Joerg

Well, like I said, I did, and the plugs fit, and it halfway works!

Basically once the pins on the side of the "huge" NUMALink 3 port are removed the "little" NUMALink2 cable will fit. And my O350s, for some reason or another, arrived with the pins in plastic baggies instead of actually screwed into the port, thus why I was able to plug my NUMALink2 cable in.

And, like I said, the L1<->L1 communication works, just not the whole transferring data part xD

Hmm, are you perhaps connecting the cable to the older XIO port designed for O2K gfx options?

I cannot see how you'll ever get it to work if you continue to use an O2K cable. Find an O3K cable.

Ian.

ShadeOfBlue wrote: Note the "decrease SGI revenue", the export laws are probably just an excuse

Given how fast commodity clusters had already become by then, probably yes.

Irony is, Russia had a better chip anyway (Erebus2K), though it floundered in the end, bought out by SUN.

Ian.

Note the "decrease SGI revenue", the export laws are probably just an excuse

No - the export laws are there any very tough, it is an issue I come up against a lot (certain S/W, H/W I/O speed etc)...

mapesdhs wrote: I read that too somewhere, but it was dumb, whoever thought it up. Paranoia gone mad.

The US has the paranoia quota for the entire world all sewed up. Communists ! Communists ! The totalitarian repressive Commnists are going to eat your children !

We need to institute IQ testing as a voter qualification.

hamei wrote: We need to institute IQ testing as a voter qualification.

Hmm... in the UK, I reckon simply being able to speak properly would be a better starting point. 8\


Anyway! Back to more cheery things. Any more upgrades planned for Iceberg?

Ian.

hamei wrote: We need to institute IQ testing as a voter qualification.

In Australia it is illegal not to vote... so all the legally brain dead are forced to vote when they otherwise wouldn't bother ... which is why we have someone who eats their own ear wax during parliament sittings (its on youtube) when not mumbling gibberish and the lead singer of the rock group "Midnight Oil" ruining, I mean running, the country (into the ground).

RR.