SGI: Hardware

A lot of people have recently been asking about (or at least implied a concern about) the power usage of their sgi systems.

I borrowed a friend's current clamp meter last weekend and measured the power usage for my two big SGI systems. Here are the results, in case anyone's interested. I'm also hoping to solicit other folks' experiences here -- hopefully the result may lead to a comprehensive wiki page on actual power usage by system and possibly component as well. I realize there are a lot of variables, but hopefully over time (and with a bit of crowd-sourcing) we can establish a fairly complete data set. If nothing else, it may at least be amusing for posterity.

I have the Onyx2 hooked up to my water heater circuit, which provides 244V in opposing phases, so the current was measured (and found equal) on both hot wires:

Onyx2 Deskside IR3: 2x RM10, 2x Node boards (8G ram), PCI card cage + 2x PCI cards (GbE & FC), 1x HDD, 1x CD
Idle: 3.8A per phase x 2 phases x 122V AC = 927W

Onyx2 Deskside IR3: same as above except only 1x Node board (4G ram)
Idle: 3.2A per phase x 2 phases x 122V AC = 780W
gfx fully utilized: 3.6A per phase x 122V AC = 878W

So it looks like:
-during gfx-intensive tasks, a 2rm IR3 uses an additional 0.8A or ~100W
-a node board (IP31) full of ram uses 1.2A or about 150W.


The 4D/240 runs on standard 120V mains:

4D/240VGX twin tower: IO2, 2x IP7, 1x MC2 + 12x2MB=24MB, std 4-board vgx, 1x VP1, IO2, 2x stackable drive modules, 3x 1/2-height HDDs, 1x tape drive:
Idle: 9.7A x 122V = 1183W

This is interesting because the power supply is only rated at 1050W. Perhaps the VGX transplant is too much for this little guy? It certainly makes a 16-gauge cable warm (so I've switched to 12).

Since there's also so been some mention the advantages/disadvantages of using 240V AC or 120V AC, it would interesting to see your Onyx2 tests repeated with the system connected to 120VAC.

I've contributed some power figures on the Origin350 and NUMAlink module already:

http://www.nekochan.net/wiki/Origin_350
http://www.nekochan.net/wiki/NUMAlink_router

I have started a wiki topic and there is a lot of information already in the forums... Was obi-wan who said, "Use the search, Luke". No, wait that might have been the force.. almost as powerful.

http://www.nekochan.net/wiki/Power_usage

Knock yourself out editing the hell out of that page...

Hi I know that on O2 HW agregator is power consumption of O2 measured. But last week I'v done measuring of my system R7000@600, 2xHDD, 1xCDRW, 2x128+4x64+2x32 RAM, AV1 board and camerra attached to it. It took 94W idling and 110W peak witch is bit more than in agregator. Measured on standard 220V wall outlet.

Power consumption of my Origin 3400 with 16 400MHz R12k CPUs and 16GB RAM (all RAM slots used):

These measurements were made when the system had only one 36GB 10k RPM hard drive. Since then, I have replaced it with two 73GB 15k RPM drives -- I should measure the power consumption again to see if it changed much. I think it's safe to say it's still under 1400W

zuluchas wrote: This is interesting because the power supply is only rated at 1050W. Perhaps the VGX transplant is too much for this little guy? It certainly makes a 16-gauge cable warm (so I've switched to 12).

The 1050W number is the combined output power (U*I) of all power modules of the PSU. You have to add the power lost in the PSU to calculate the power drawn from the net. I believe they are quite efficient especially considering their age, but even (let's say) 80% efficiency means it generates 200W heat inside the PSU. There's a fan in it for a good reason!

There is another rating (1500W) for the max input power the that the primary stage can handle.

These two numbers for the same PSU, combined with the fact that you have to disassemble quite a few parts before you can even read the label, has led to endless discussion threads in comp.sys.sgi about "RealityEngine capable" PSUs and which system configurations would need those. In the real world, my 4D/440VGX consumes significantly more power than my Crimson RE.

At least, that's how it is for a single tower PowerSeries. Here's the label for the PSU of one:

I've been collating some results, but need to redo a few of them as the meter was in the wrong mode in some cases (VA instead of Watts).

http://www.sgidepot.co.uk/sgipowerusage.html

Indigo2 IMPACT is the big juice chomper, by a mile.

I would test my quad-1GHz Tezro, but it currently lacks fans/VRMs on the CPU board. Had to move them when I bought a board
on eBay from which these items had been removed (nuts!).

Ian.

quite unexpected results, especially:

On O2, if the resolution is reduced from 1280x1024 to 640x480, the power consumption drops by 0.01A.

thanks for your 'services' once again

A 1% drop in power consumption is below the noise floor... Interesting to note, but not a meaningful statistic...

The Keeper wrote: A 1% drop in power consumption is below the noise floor... Interesting to note, but not a meaningful statistic...

Wrong, it was completely repeatable.

Ian.

mapesdhs wrote: http://www.sgidepot.co.uk/sgipowerusage.html

Ian, you're welcome to add my results (above) to your list. I'd like to see some more figures for "big iron" SGI systems eventually. The next time I get a hold of a current clamp I'll do some more testing on the systems I have to hand as well.

Cheers!

Wow - O2 has quite decent performance/watt - too bad the total performance was always so low

A 12W+ standby drain out of I2 is pretty annoying (I'd be flipping switches on a powerstrip on that for sure), but nothing like other systems of its vintage - my AlphaServer4/266 tends to leak over 80(!!!!!)W when off (somehow).

> Wow - O2 has quite decent performance/watt - too bad the total performance was always so low

Indeed, if one considers a typical render job, O2 takes longer but uses less power overall, eg. with
an R12K/400 it's only a bit slower than an Octane2 400/V6 but uses barely 1/3rd the power.

Even for the complex Alias test, Octane is 2X faster but (unless it's headless) uses 3X the power. Only
good dual-CPU Octanes, or other systems such as Fuel, start to win overall. On the other hand, not
everyone is happy leaving an appliance on overnight doing something.


> - my AlphaServer4/266 tends to leak over 80(!!!!!)W when off (somehow).

Ouch!

Ian.

mapesdhs wrote: On the other hand, not everyone is happy leaving an appliance on overnight doing something.

Oh yeah, I almost forgot ... you guys have Joe Lucas refrigerators over there in England, don'tcha ?