Re: 10x?! (was guaranteeing Friendlyness)

From: BillK (
Date: Mon Dec 05 2005 - 16:10:35 MST

On 12/5/05, J. Andrew Rogers wrote:
> On Dec 5, 2005, at 9:33 AM, BillK wrote:
> > <>
> > 96 processor deskside workstation, peak performance of 230 gigaflops,
> > running off a single electric plug.
> >
> > Cost around 100,000 USD.
> > But big savings on data centre costs, power, space, rackmounts, air
> > conditioning, management, etc.
> Unfortunately, this product emphasizes power efficiency above all
> else, including price performance and absolute performance. You
> could put a low-power Opteron cluster in a single rack for similar
> amounts of money and get far better performance for maybe twice the
> power budget. That difference in peak power consumption (retail)
> including cooling etc, is on the order of $1000-2000 per year. In
> other words, negligible.
> But since the Opteron cluster will be so much more performant for
> many (most?) workloads, you'll likely get the same amount of work
> done per watt, plus the implied time savings. For codes that heavily
> exercise the memory rather than spinning in cache, and many AI codes
> seem to be of this nature, Orion would be a poor match.

Well, yes, but nothing's perfect is it? :)

The specs say you can have up to 2GB DDR memory per processor. 192GB max.
But the point of this box is that it is an off-the-shelf Linux cluster
which slides under your desk. That's the installation done.

Basically, my attitude is that tinkering with the hardware and the OS
is an unnecessary distraction from the business in hand. That's why
businesses have a dungeon full of geeks totally separate from the
company enterprise.

Rolling your own can get increased performance. And you can keep
fiddling with it to improve matters. If you have techie time available
but little real cash then you have no choice. But you run into a load
of other indirect costs. Your time (or a techie's time) building and
tinkering with the gear is a cost. You might need extra air
conditioning or a bigger power supply to the building. Your
environment degrades with racks and cables - but it does look more
like a real university AI lab. :)
Your insurance company might not cover DIY kit. You might run into
Health and Safety regulation problems, or Fire Safety regulations
problems. You have less of a guarantee if bits break. The % uptime is
probably less because of fiddling with the kit. It might even get
accidental damage while being 'improved'.

It is up to each project how worthwhile it is to spend time and money
on developing their own unique Linux cluster hardware and software.
And I appreciate that playing with the kit and tuning the OS is a lot
of fun for some people. But developing AI code is a hard enough
project on its own, without all that extra hassle.

If you already have a small supercomputer under your desk I doubt that
it really matters all that much that you could get even faster kit. At
least until your project development approaches the limits of what you
already have installed. It will all be replaced in a year or two


This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:54 MDT