The mystery behind secretive server startup SeaMicro is dispelled today as the venture-backed maker of what it has been calling "data center appliances" unveils its first product: the SM10000, a server cluster comprised of 512 of Intel's Atom processors with a built-in, virtualized network fabric for the servers. The SM10000 …
You want atom supers?
You have no taste. Try loongsong now. Or back when it came out, it would've been amazingly cool to do this with pmc sierra's rm9000. Or the nec VR5000 series. Looking forward, a multi-core low-power >1GHz ARM.
Let's face it: x86 is so last century. atom is wholly uninspired. Transmeta brought some truly interesting tech to the table. Intel did its level best to kill it. Then it gave us itanic. Which, by the way, cost us two better performing architectures. Oh, and atom. Mostly to kill off other initiatives. Using intel while crying "innovation" is a bit like micros~1 crowing "invention" with the "micros~1[tm][r][c] surface[tm][r][c]" when a couple german blokes had already done the same thing in a cellar under Berlin.
Yes, everybody does it, everbody expects micros~1 to pull the same "look what we have, it's <something someone else already did better>!" trick over and over again. That's what they call "research". It just isn't very exciting unless you really are that stupid, shortsighted, uninformed, that much of a poser, all of the above.
If this is the best we do, we are getting what we deserve. Babbage would not able rightly to apprehend the kind of confusion of ideas that provoked this state of affairs.
Imitation is the sincerest form of flattery
Atom is Intel's take on Via. It imitates most of its features (some of them badly) and fails to deliver on a lot of counts.
However, while imitation is the sincerest form of flattery, Via should only blame itself for failing to exploit its 7 year lead in low power / ITX systems.
The intel anti-fanboi speaks. Think you missed the point of the article. ITs not that it used Atom its that it ties lots of cheap processors together in a clever dense package.
When you stop being a whinetard about x86 dominance of a huge part of the general computing space maybe your opinion will count for somthing but I wont hold my breath.
Until them try hitting yourself with a cluestick - the rest of us would help but we have better things to do.
Well, with no IONs in that atomic cluster, it won't play Crysis but what about the ips and flops?
...the usual commentard question: "But will it run *******?" (insert name of whatever silly game currently has insane HW demands)
will it blend!?
and you shall receive
will it run crysis 2, metro2033 and left 4 dead 2 AT THE SAME TIME?
HA! bet u didn't expect that :D
My chip of choice...
I love the very much under-rated Zii Labs chips such as the ZMS-05 and ZMS-08.
Low power, high math and easily managed.
I'll take one of those for some serious number crunching at low power!
It sounds a lot like those are individual systems, so I have to license and/or maintain 512 copies of the OS?
You almost have a point, except the OS isn't installed 512 times, its a single install on the cluster, and seriously this is an HPC cluster for all intents and purposes, 94% of them will run some random flavor of Linux 4% will run BSD and the ones that do get windows will be dual boot because the owners will have too many reservations about the stability of windows at that scale...
Well, you could do so but if you're talking on this kind of scale, it is generally easier to scale up Linux type apps than it is Windows based...
You're not going to be talking about 512 OS instances, probably nearer 64 for the mainboards (since it seems to be a 64x8 setup, rather than 512 totally independent cores), which means 64 licences that can handle 8 processors (which rules out the lower end of Windows licences anyway)
But seriously, you're gonna be running this bad boy for database stuff most likely, or webserverish nonsense, which screams *nix anyway.
RE: stability of Windoze at that scale
What a sight that would be when a box containing 512 instances of Windoze completely crashes and burns.
I want this.
So many workloads have no need of anything more powerful than an Atom to get the job done. This is physicalisation done right.
Require now plox.
Very neat indeed...
Me like it...
Hare & Tortoise
The Hare clearly got the better press these days. No wonder, given Oracle recently had Sun Microsystems for Tortoise soup.
The whole ad^Warticle read pretty much like Sun's marketing ooze about "CoolThreads technology" and why you need hundreds of twice-powerless cores for web2 workloads. T5440, anyone ? And when again was the age of the blade servers ?
Must fight the deja-vu. Box looks cool. In ten years, we'll be chilling beer inside.
Pick your evil (ed: please, we need an icon with Steve Jobs' face, Larry Ellison's nose and Bill Gates' glasses; add comrade Stalin's beard for good measure).
Still not cost effective...
In theory we should be an ideal customer for this sort of thing, but I would not buy one for the following reasons:
1) They have got the RAM-CPU ratio wrong. RAM is by far the most important resource these days, especially when using mass virtualisation. Needs twice the amount of RAM per CPU ideally (then it would be better than what we can currently buy).
2) I can buy 43 quad-core 2.5GHz xeon servers (probably a little more CPU than 512 atoms), with 24GB RAM each (1TB total) for half that price tag (£50k).
3) The 43 pizza boxes would use about 5kW. Over a 3 year lifetime, in a modestly efficient data centre the 3kW would cost me an extra £13,500, so the power saving is dwarfed by the additional cost.
4) Their "magic" clustering and resource partitioning stuff does not sound any more impressive than what we currently use to glue together our large clusters and virtualised grid.
Oh, and before someone goes "but space is expensive" - space is only expensive if you're in the tiny minority of people who really, really must be inside the M25. Otherwise, power is the limitier.
Good idea to use standard hardware interfaces (PCI etc) though. Double the RAM or halve the price and this would be useful.
The article mentions Atom Z530 and then x64 - Z530 doesn't support x64 - only x86.
@Still not cost effective..
Depends, that's list price, at list the Xeon boxes would cost the same.
Remember that whatever you pay for power-in you pay 3-5x as much for cooling to get power-out, more in Texas or New York in summer.
Space IS limited if your server room is full, this is a lot quicker/cheaper than planning permission to build another server room next door.
how what short memory's are, remember the PowerPC Rapport Kilocore1025 cores
while its impressive, finally a reasonable low power inclusive Generic PCB on a generic IO card shot
did People forget about the 'Rapport Kilocore1025' 1024 cores +one ,and the Kilocore256, which comes with 256 processing elements and provides a performance of "25 GB operations/second at well under a single watt of power,"
the draw back way back then in 2006 was OC these "1024 (that is one thousand and twenty four) 8-bit "processing elements" on a single and - according to IBM - low-cost die."
so a single die rather than a larger so called credit card size for these 1024 and 256 cores but only installed 8 bit elements rather than 32/64 bit elements internally, but that was a fully functioning and available die 4 years plus... so adding 32/64 bit would not be a problem id hope in combination with today's die shinks etc....
OK, I too want one, just to play with, not really sure what I'd do with it but it's too darn interesting to ignore! I can't remember if the Z5x0 Atoms are limited to x86 and not x64 as Ender posts, but that's not too much of a problem unless you want to play with a large memoryspace per core. I'd be interested in seeing how these systems could run as a virtualised desktop server(s), with maybe eight desktop clients per card, running something like RedHat's Enterprise Virtualisation for Desktops. In theory, you could cram a whole small bussiness's or medium-sized office's desktop estate into one or more of these jobbies. Only power users really need a dual-core Athlon or Duo chip PC, most would be quite well served by even a single-core Atom, let alone a virtualised share of eight.
I'm also curious as to whether the tech could be used for an Atom-based offload engine card. If an eight-Atom "server card" in the current design needs two PCIe interconnects, could they make a four-Atom card with one PCIe interconnect that could slot into an ordinary server PCIe slot (or a PC one even)? Suppose I have a two-socket x64 2U server such as the DL380 or PowerEdge R805, and I put in four of these cards that co-operate over the PCIe bus and can take advantage of the server's memory, internal disks and interfaces.... Or I could use the cards for dedicated servers inside the main server chassis, such as using one four-Atom card as a NAS server (hp and CISCO have single-Atom stand-alone NAS boxes) with a couple of onboard Gb ports and using the main server's internal disks or SAN connects as the staorage.
Maybe the big players will sit up and take note and start offering Atom-based blades, offload cards and plug-in-server PCIe cards. I suppose it all depends on whether there really is much of a market for them.