Linked by Eugenia Loli on Mon 11th Nov 2002 19:56 UTC
SGI and IRIX A new supercomputer from SGI packs far more computing power per square inch than do competing machines, a breakthrough the company hopes will help it flourish in what it sees as a renaissance in supercomputing.
Order by: Score:
Lets see how well it does on SETI@HOME
by al pettit on Mon 11th Nov 2002 21:00 UTC

They should re-compile seti@home on it and see how many units are done in a day

???
by jon on Mon 11th Nov 2002 21:35 UTC

64MB??

I assume they ment 64GB.... unless MB stands for mega^2 bytes :-).


Jon

?
by JJ on Mon 11th Nov 2002 22:26 UTC

Interesting but there must be a few mistakes.

Now if $2.9 buys 128 cpus, thats $24k per cpu node, I am assuming these nodes are mostly cpu mem & links.

Although x86 doesn't make the ideal super computer node, too much heat, space & not the best idea for such a thing, an Athlon node can easily be built for $240, about 100x cheaper, there must be something wrong here!


my 2c

RE: ?
by jon on Mon 11th Nov 2002 22:33 UTC

i think that the cpus are MIPS not x86, but i could be wrong

RE:?
by JJ on Mon 11th Nov 2002 23:07 UTC

I wasn't suggesting that they were x86. Article clearly says MIPS & MIPS corp is/was an SGI subsiduary so x86 would get short thrift there. However $24k a cpu is still very high, I would guess the article is inaccurate or the modules are quads or something. Will have to go to SGI for the real info!

Also the last time I looked, MIPS like PPC was barely at the 1GHz level so a single MIPS chip pc would not now compete against x86 on price or speed (anybody remember NT for MIPS), but its still a fine cpu architecture, and the power 17W does give it a huge power/power (mips/watts) ratio over most everything else.

re: ?
by Kevin on Tue 12th Nov 2002 00:33 UTC

Although x86 doesn't make the ideal super computer node, too much heat, space & not the best idea for such a thing, an Athlon node can easily be built for $240, about 100x cheaper, there must be something wrong here!

From the article:

"[...]density increases performance by reducing the distance that signals have to travel between the processors, memory and other subsystems. The subsets of most large supercomputers are linked through cables. Because signals can't be accelerated, the best improvement lies in pushing the machines closer together."

re:?
by JJ on Tue 12th Nov 2002 04:17 UTC

I am sure that SGI packs the MIPS chips as close as it can probably many times closer than x86 would allow. Also sure they could probably get something like Hypertransport in there to exchange data between cpus, where this still isn't here yet for el cheapo x86 on a large scale. But still, I think I might take a 100 Athlons with 100/1G networking v a single MIPS node for same $, but it sure would be a hot place to be. Clearly this machine is vastly more closely coupled than packing a bunch of x86 boxes together. Pity the article gives no real detail.

Re: Good conversation on Slashdot
by Rayiner Hashem on Tue 12th Nov 2002 05:54 UTC

Seriously. They mention the advantages of this machine, particularly the fact that the internode latency is far smaller than in a cluster machine. For non-parallizable problems, a single system image is much faster than a cluster. The 24K per machine cost seems much better if you consider that a lot of the cost of the machine is in the high-end interconnect. Hypertransport is actually a very interesting beast. It's derived from the massive Alpha 21364 interconnect, so it has a high-performance computing heritage. It has a higher per link bandwidth than NUMALink (SGI's interconnect) and it's 70ns per hop latency is probably comparable (worst case NUMALink is around 700ns, I don't feel like doing the match to see what that is per hop). The only question is whether hypertransport can scale that high.

MIPS vs. x86 for supercomputers
by matt on Tue 12th Nov 2002 09:55 UTC

plus there arent any x86 chips which run > 8-way are there? certainly no more than 32-way, yet this machine is supposed to be massively scalable cc-NUMA.

x86 chips have pretty poor links off-chip as well, at the moment they cant handle the i/o required for massive supercomputing. simply put: they cant shovel data from memory to the chip fast enough.

certainly x86 is a good choice for a cheap, high-latency cluster, but not for a supercomputer.

Athlon Supercomputer
by Hank on Tue 12th Nov 2002 13:23 UTC

an Athlon node can easily be built for $240, about 100x cheaper, there must be something wrong here!


Is this a joke? First of all, you can't make a reasonable Athlon node for that price. You will want a top end Athlon and a gigabit ethernet card, or a series of 100-Base-T ethernet cards, and 512MB of RAM. You won't get this, and a power supply/case and POS hard drive for $240. It won't cost you $24,000 either. However, you are also not getting the high speed interconnect or a system of 128 processors that can fit into a single rack.

As I've said ad nauseum in here when the parallel supercomputer topics come up, PC clusters work very well for many applications. Some applications still require and take signficant advantage of the big iron supercomputer technology.

Don't try to fit the square peg in the round hole.

Hank

AMD in "Real" supercomputer
by Hank on Tue 12th Nov 2002 14:44 UTC

Sandia's new beast, based off COTS hardware:

http://www.sandia.gov/news-center/news-releases/2002/comp-soft-math...

Without reading the article, the per-node cost (assuming one processor per node) is $5625. Not bad!

Think of a
by Anonymous on Tue 12th Nov 2002 17:16 UTC

beowulf custer of those babies