AMD shows off dual-R600 "teraflop-in-a-box" system

NVIDIA isn't the only one putting its massive amounts of GPU horsepower behind more mundane -- and potentially more lucrative -- tasks than pretty 3D gaming graphics. AMD just showcased its Accelerated Computing platform, which ties an AMD Opteron dual-core chip to a pair of AMD R600 Stream Processors for more than a teraflop of combined performance. AMD's not only proud of the basic muscle on display, but the achievements of its Accelerated Computing platform getting all that beef to work together. But really, all we want to know is if it can run Doom.

[Via DailyTech]

[Benson Leung @ Mar 2nd 2007 12:57PM]

Technically, since the two stream processors are just two R600 graphics cards, they can and do run Doom.

[Rat Bastard @ Mar 2nd 2007 12:57PM]

"Teraflop-in-a-box" was the funniest thing on SNL in years, I heart Justin Timberlake.

Oh, wait a sec...

[beanz @ Mar 2nd 2007 1:31PM]

better question is "How well can it play Doom?" and "When can i buy it?"

[Brennan @ Mar 2nd 2007 1:33PM]

well i already know AMD, or DAAMNIT, is making a dual GPU card using 2 R600s, similar to the 7950 GX2.

my understanding of teraflops or flops is confusing to me, so it sounds good, but i dont understand what it means, heck i even looked on wikipedia.org for the definition but i still dont understand it. Can someone elaborate in a sense of understanding plz? I appreciate it.

im planning on getting a pre-OCed X2900 XTX which sounds better in my opinion.

[Mike @ Mar 2nd 2007 2:16PM]

A Flop is a floating point operation, a calculation using numbers with a moveable point, such as 3.234, 0.008 or 125.456 . They are numbers where the decimal (binary, technically) point can be at different places. For example, adding 1.22 to 100.001, this is a floating point calculation, the numbers have the points in different places.
other types of calculations are integer (whole number e.g 10, 211 or 1002) or fixed point (1.20, 12.32 , the point is always in the same position)
Floating point calculations are the most useful for making complex calcultions and realistic simulations. Modern games rely on them!

A teraFlop is one trillion Floating point operations per second!

[Brennan @ Mar 2nd 2007 1:39PM]

from what ive heard, i suppose to come out possibly this year, Q3-Q4 2007 after the launch or so of the high-end X2900 XTX.

[Brennan @ Mar 2nd 2007 1:39PM]

from what ive heard, it suppose to come out possibly this year, Q3-Q4 2007 after the launch or so of the high-end X2900 XTX.

[ethana2 @ Mar 2nd 2007 2:22PM]

Flops are just a measurement of processing power related to large amounts of math. If you're folding proteins, doing real time physics, or generating high resolution bitmaps from 3d images, the flops are all important. Tera- of course means "Trillion".

What I want to know is how many gigaflops per watt it does, whether or not it's Linux compatible, and how many idle power savers are used, like dynamic voltage, clock throttling, and anti-leak clock gating.

No way I'm going to waste money on Windows-only hardware. And a GPU that can only do graphics is ALMOST as pointless as a console that can only play games. Stream co processors, though are the epitome of my PC ideals manifest.

[Blake Kachman @ Mar 2nd 2007 2:27PM]

Now ditch the opterons and run that same configuration with a Quad Core2 and you'll really be in buisness.

[Chris @ Mar 2nd 2007 2:36PM]

My reading of that says that they are using the r600 core not for graphics but as a math co-processor to the opteron. back in the day adding a math co-processor to a 386 or early 486 gave a huge performance increase at the same clock speed. IIRC, the 486DX was the first intel chip to have that co-processor built in. In modern CPU that co-processor has evolved to SSE and 3DNow. however, GPU's are build specifically to do floating point calculations very quickly and using one as a co-processor is a great idea, if you can handle the heat dissipation and power requirements which should be no problem in a server.

what does this mean to the average user? nothing. your graphics chip is already doing all the FP calcs you need done. In the enterprise though it means that HTTPS can be processed faster and thus a single server can handle more users at once, and possibly would help with database indexing as well making the database respond faster, so I guess it could help users in this way. it will also make current clusters and super computers that much faster too.

[Kiteless @ Mar 2nd 2007 4:47PM]

Hate to correct you chris, but floating point math has nothing to do with web servers. Stream Computing is going to be great for schools. Imagine being able to do fluid dynamics in real time on a $3000 dollar machine instead of a 3million dollar super computer. This is also great for weather prediction, physics calculations like bridge construction (testing stress levels), hydro dymanics, things like that. I think stream computing is going to change the face of super computing as we know it.

- K

[James @ Mar 2nd 2007 3:22PM]

I'd really like to see a list somewhere of what software (apps and/or games) are actually coded such that these extra processors will be used -- I know a big problem with dual core/dual processor systems is that for years, the software industry didn't have a firm grasp on good multithreading, and as a result much of today's legacy apps don't see much benefit from adding an extra core or two. Won't this AMD system suffer from the same dilemma?

[Castle @ Mar 2nd 2007 3:42PM]

single-point precision teraflop is which world's away from double-point precision that real super computers used. Fine for games, but not really a "teraflop" in the sense that usually hear about them.

Good for viral marketing hype though. . .

[WhIteSidE @ Mar 2nd 2007 5:46PM]

I am sure that someone will have an mpi compiler for these pretty much as soon as they come out (or soon after).

Furthermore, even if this is not the case, I would imagine that it will not be too hard to write code for them, it will just be an adjustment for people who never do multiprocessor operations. That having been said, as a comp sci / physics student, this would be so amazing.

I am doing research with fractals, and sometimes have to wait days to get processing power on a "super computer" (a beowulf cluster). A couple of these rigs could replace dozens of opteron 275s for a tiny fraction of the price. And that's great for me. I sometimes run sims with several thousand trillion floating point ops or more, so a single system like this could run a sim like that in only a couple hours.

Way cool.