[Beowulf] AMD Roadrunner open compute motherboard

Lux, Jim (337C) james.p.lux at jpl.nasa.gov
Wed Jan 16 17:49:46 PST 2013


What I found more interesting was the origins of this... a software service vendor (Facebook) sponsoring the development of a hardware platform for their use.  Interesting.

I don't know how useful the "open" aspect is.  The startup costs to make a batch of mobos are fairly high, so it's not a "small volume" kind of proposition, and I'm sure that for any volume at all, you could get the reference design from the processor mfr and start there.  For all I know, the reference design isn't "closed source".

I thoroughly agree with Bill on the fans and such.  Maybe in their target application, they've got external fans or ducting.  They've already moved away from the 19" RETMA rack.. Maybe they're going to the Lux Baking Sheet packaging model.  

16.5 x 16" form factor would fit nicely on the end of a 26x18" sheet pan.  The inside dimensions of a standard sheet pan are about 16.5" minimum (depending on how much the sides slope in and how big a bead there is.  That would leave you about 8" for your disk drives.

The Facebook Open Rack V1 is 23.6" wide and 42" deep.   A big sheet pan rack is 23-26"W and 30-39" deep .   Facebook open rack has adjustable ledges upon which the computers sit. Some bakery racks have adjustable ledges upon which the sheet pans sit.

  Coincidence?  I think not.<grin>

If we find out that they are using double stick foam tape to hold the motherboard down instead of fasteners, then we'll know.. 
you read it here first on the Beowulf list.. 
(actually, it says "It is completely screw-less," and the pdf of the chassis says it is 18.9x23.38 inches.. a bit different than a standard sheet pan)

The airflow spec says 12-106 CFM at the server level, and 1116-9324 CFM for  a whole rack with 90 servers, stacked 30 high and 3 wide.  
http://www.opencompute.org/wp/wp-content/uploads/2013/01/Open_Compute_Project_Server_Chassis_and_Triplet_v1.0.pdf


But really, for a high density system, separating the fans from the electronics, when you're already going to have 20 or so shelves in the rack just makes sense.  You're never going to try and pull out one system while leaving the rest running.  In these kinds of massive shared load datacenter apps, why not just let the systems die, and accept the slight degradation as your shipping container with 2000 processors slowly turns into a 1999, then a 1998, then a 1997 processor machine.  Maybe once every couple of months, you send someone through to shut down a whole rack and replace the failed system within it, and then bring it up again.




Jim Lux

-----Original Message-----
From: beowulf-bounces at beowulf.org [mailto:beowulf-bounces at beowulf.org] On Behalf Of Bill Broadley
Sent: Wednesday, January 16, 2013 12:52 PM
To: beowulf at beowulf.org
Subject: Re: [Beowulf] AMD Roadrunner open compute motherboard

On 01/16/2013 10:20 AM, Hearns, John wrote:
> http://www.theregister.co.uk/2013/01/16/amd_roadrunner_open_compute_mo
> therboard/

The pictured 1U has what looks like 12 15k RPM fans (not including the power supplies).  Or 6 double fans if you prefer.  In my experience those fans burn an impressive amount of power (40-60 watts), make an impressive amount of noise, and introduce substantial vibration into the chassis.  That might be worth it, but they don't actually move much air.

I'd much rather have a quad node 2U chassis (with 2U fans).  When compared to the picture 1U you get:
* 1/4th the power cables per node
* less chassis per node
* efficient fans/airflow
* better power efficiency
* a price that's often less than the four 1 U nodes.

_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list