96 Processors Under Your Desktop

Roland Piquepaille writes "A small Santa Clara-based company, Orion Multisystems, today unveils a new concept in computing, 'cluster workstations.' In October, you'll be able to choose between a 12-processor unit for less than $10,000 or a 96-processor system for less than $100,000. These new systems are powered by Efficeon processor from Transmeta and are running Fedora Linux version 2.6.6. Apparently, this new company has friends in the industry. You already can read articles in CNET News.com ("A renaissance for the workstation?"), the New York Times ("A PC That Packs Real Power, and All Just for Me," free registration, permanent link) and the Wall Street Journal ("Orion Sees Gold in Moribund Workstations," paid registration). The company is targeting engineers, life scientists and movie animators. It's too early to know if the company can be successful, but I would certainly have to get one of these systems under my desk. In this overview, I've picked the essential details from the three stories mentioned above."

strange

[ZenBased]

why not get a huge server where more users can benefit from the processing powers? and what kind of videocard does this baby pack? that must give some great doom3 performance :)

Re:Cooling?

[pedestrian crossing]

That was my first thought. I guess that's the point of going with Transmeta...

Fedora 2.6.6?

[Halo-]

I hate being this guy, but this is a big pet peeve.

Fedora currently is either Core 1 or Core 2. 2.6.6 is a kernel version number.

Kernel version != Distribution

Saying "Fedora 2.6.6" is like saying a car is a "Ford 2.4 liter".

Seems Very steep

[Crashmarik]

The pricing seems quite steep. 800/cpu for 12 configuration, 1000+/cpu for the 96 cpu configuration. I can see why they have friends in the industry the prospect of selling 10 to 100 times the equipment per seat must have marketing departments salivating.

If your'e going to spend that kind of money though theres alllready solutions that will provide that level of processing cheaper.

There is also the utilization isssue, programming tasks hardly require 96 processors except on compile and link. You don't need 96 processors to wait for a keystroke. The same holds true in applications. You don't need cpu's waiting for a user to decide what to do.

Re:For a moment...

[dreamt]

I woiuld guess it is more a physicical issue. Processors are probalby arranged 3x4, and can fit that way into a 2 foot wide case to fit onto an "average" desk. Stacking these boards 8 high gets to a height that can fit under a desk.

Re:Sounds nice, but

[Xocet_00]

If I had to guess I'd say that the on-board communication would be switched, such that the chips can talk one to one at 1Gbit.

For chips on different boards to talk though they would need to squeeze their traffic down the same line as all the other chips trying to talk board to board. Hence the higher bandwidth?

Just a guess.

Re:For a moment...

[skaffen42]

Could be cost. 10K and 100K are nice round numbers.

I've realized that most strange tech descisions can usually be traced to some guy in sales...

Re:Sounds nice, but

[Dutch_Cap]

"Wouldn't same board communication be more frequent, hence needing the faster connection?"

I guess it depends how you look at things. On the same board you have one processor talking to one other processor. Between boards, however, you have up to twelve processors talking to up to twelve other processors. So to me it makes sense to me to have more bandwidth between boards than internally on a board.

Interconnect & SSI Required

[Anonymous Coward]

The interconnect is tcp/ip over ethernet, just think of the the overhead it generates on the poor transmeta CPU. Actually, quite strange given that each efficieon has its own Hyper Transport Bus they could have been much clever about it and on the way also use Single System Image OS - maybe Mosix. As I see it there is not much usability for the staff, unless you like noice and heat beneath your table...

Price/Perfomance

[Anonymous Coward]

So what's the benefit? How does this compare to a dual opeteron or something similar? If a dual opteron can do something like 12G flops for $2.5K, what can this thing do with 12 processors at $10K...?

Re:For a moment...

[Peter La Casse]

Seriously, why 96? Why not 64 or 128?

12 processors fit on one board, and 8 boards fit into the chassis they chose.

Reliability

[lachlan76]

This wouldn't be as reliable as having 96/48/24/12 computers with 1/2/4/8 processors each, which would be important for things like movie animation.

And besides for movies, we already know to just fit as many Opterons in a rack as possible. What advantage does this have (except for cost)?

Re:Seems Very steep

[AtomicBomb]

The CNet articles explains why the 12-way version is cheaper on a per CPU basis. The chips of the smaller 12CPU version are all mounted on single board and connected with 1Gb ethernet... While the 96-way version connects eigth 12-way CPU board with 10Gb ethernet... The high speed communication may make the 96-way system more expensive to start with.

There is also the utilization isssue, programming tasks hardly require 96 processors except on compile and link.

However, computer users are more than just programmers and/or IT people... Many scientific applications and animations require parallel computing... Basically, the more the better for them. They can use up any resources you throw to them. To them, the $800/CPU pricetag is not that expensive... A Sun 8-CPU machine costs them way more than $10k... A dual Xeon Dell machine with 8GB RAM/ 800GB HDD cost more than $7200...

Re:Price/Perfomance

[Jeremy Erwin]

The whole point of this marketing exercise was to bring cluster computers out of the glass walled (and blinkinlight laden) server room, and to the desktop, on the assumption that the bureaucracy of node allocation gets in the way.

With this setup, a mathematician can get a flash of inspiration, fire up grid Mathematica, and have 12 processors testing her hypothesis in a matter of seconds. A biologist can run BLAST without having to worry about whether his colleagues might be hogging the computational resources.

Essentially, it's a very expensive "personal cluster" machine,

I agree completely. HEMOS, listen to me...

[Ayanami Rei]

Can the editors PLEASE STOP POSTING PRESS RELEASE COPY?!?!

If you're going to post a story announcing a product or discovery, at least link to a weblog or site that actually has a little commentary on the subject, or the original site itself.

Roland "Fuckyfacey" Piquepaille is neither of these.

Thanks.

Re:Seems Very steep

[CommieOverlord]

The 12-CPU system here scored 18GFlops sustained (using Linpack), 36GFlops peak. That's 1.5/3 GFlops. The CPUs are 1.4GHz, so that's 2.15flops/cycle.

By contrast, Xeons are 2flops/cycles. So a 3GHz Xeon can achieve 6GHz peak. Given the efficiency of the processor, and inefficiences from parallelizing,expected performance is probably around 3-4 GFlops.
A dual 3GHz server runs about $2500, so that means 8 Xeons for about the $10K price of the Orion. So that's 24-32Gflops expected, 48Gflops peak.

Based strictly on performance, the Xeons have the edge, however

1.Space - A a company doing rendering could stick a couple of these under desks instead of allocating an entire room for computing (space costs money).

2.Electrical - The Orion uses the same amount of power as just one dual-Xeon node. The extra nodes will cost about $400-500 in electricity each year. Ammortized across a 3-year system lifespan that's $1500.

3.A/C - The Orion puts off as a much heat as a desktop. No need to get massive (expensive) A/C units to pump tons of cold air into a server room.

Re:Fedora what?

[decepetion]

Because with RHEL, you have to pay "PER PROCESSOR" licensing.

96 CPUs x $$$ == $$$$$$$ (you get the idea)

Re:Better solution...

[Anonymous Coward]

And your solution would used alot more power, generate alot more heat, and wouldn't fit neatly on my desk.

Cost is the issue and an important one

[theolein]

Being able to have one of these in a standard office that doesn't need a server room setup is the major factor, as I see it. It means that the company saves on racks, rackspace, the cost of electricity (which is going up as you may have noticed by the price of oil being over $40 per barrel), administrator costs (his salary alone will be worth it) and air conditioning for the server room.

Think of a small CG effects movie company (say around 5 to 10 employees): They want to be able to render their CG movie frames as quickly as possible in order to save money, but don't have the capital to buy and maintain an expensive server room setup along with the requisite admin type to look after it. These machines provide and effective alternative: 1 or 2 of the 12 CPU machines for the guys who do the rendering of small snippets and 1 96 CPU unit as an office renderer for resource expensive renderers.

Re:Whee!

[0bjectiv3]

We use it all the time to compare our DNA products to all known Gnomes. It takes like 30 seconds.

Clearly, you don't know enough Gnomes [wikipedia.org].

Re:Why not SMP??? This is just a-cluster-in-a-box!

[kscguru]

I can't find anything specifically to this effect, but my guess is that the Efficeon simply doesn't support SMP (they sure don't claim SMP support anywhere). Which means: no cache coherency hardware support and no multiprocessor synchronization primitives in hardware, making shared memory multiprocessing impossible.

Since the chip thus can't be used in a shared-memory configuration, a cluster is ideal. (And frankly, a better use of resources than cramming SMP chips like modern x86 into single-processor boxes and clustering those like the latest Linux-cluster-of-the-month).

Honestly - the system is brilliant. With one catch: the paraniod side of me sees some purchasing department comparing one of these, and a similar Sun workstation, then buying this one because "it has more CPU power". Yes ... if you run an app that works over a cluster. If it's a shared memory app, the Sun/SGI/IBM/whoever workstation will leave this one in the dust.

I see one very immediate use for these machines: developing cluster applications! Apps that run on 500-way clusters really do need to be tested on smaller clusters before they consume expensive cluster time - this machine is PERFECT for such testing! Or at least, far better than anything else out there...