NVIDIA Tesla 20-series GPUs promise to dramatically cut supercomputing costs
By Donald Melanson 
posted November 16th 2009 10:29PM
posted November 16th 2009 10:29PM
Sure, you've been hearing NVIDIA toss around names like CUDA, Fermi and Tesla for what seems like ages now, but we're guessing this is the sort of thing that'll get most folks to really take notice: a promise to cut supercomputing costs by a factor of ten. That rather impressive feat comes courtesy of the company's new Tesla 20-series GPUs, which come in the form of both single GPU PCI-Express Gen-2 cards and full-fledged GPU computing systems, and promise a whole host of cost-saving benefits for everything from ray tracing to 3D cloud computing to data analytics. Of course we are still talking about "cheap" in supercomputing terms -- look for these to run between $2,499 and $18,995 when they roll out sometime in the second quarter of 2010.
NVIDIA!!!!!...ok... im bored now.
Look at all those migs and megs!
Sorry, but I have to say this. Why couldn't manufacturers release cards that look like this (reference cooler design)? I mean, it's simple and minimalistic, but it looks really nice and clean. Yes, some manufacturers churn out good looking ones, but others decorate it too much with ugly designs.
Am I the only one who thinks the reference design can be better, at least sometimes?
because the pimple faced customers who fork over the cash several times a year prefer gaudy flashy designs and only when they grow up do they realize the truth about their horrid taste. really everyone goes through this phase. right?
I'm sure that the 99.9% of the population who don't have plexiglass PC cases (myself included) don't really care what their PC components look like.
What you don' like fans glowing blue inside you machine? how about a massive array of bright orange cooling plastic pipes? or giant led displays covered in fake chrome in a $300 gigantic thermaltake case?
Pretty cool to see GPUs getting used so extensively. Hopefully this translates to more applications which take advantage of the GPU's processing power for average end-users. Could be very useful for business applications, anything with number crunching or other intense calculations.
Modern businesses don't need such large numbers crunched anymore I hear.
@wwhat
Anything couldn't be further from the truth. Simulations be it for mechanical engineer's FEA analysis, or electrical engineers with RF propagation/DSP, or ISR platforms with image stabilization/geolocation/object tracking, etc. a LOT of computing power is needed. However, with these boards, one doesn't need to have access to a grid for their analysis.
@SmittyMcSmith
Psst smithy, I think he was having a jab at the recession.
they really narrowed the range of potential prices
[/sarcasm]
yeah that's what i thought lol!
they aren't potential prices, it's a range for a series of cards
It's probably like $2,499 for one card and $18,995 for a Tesla computer which has multiple GPUs.
They have cards with several GPU's and either 3 or 6GB of RAM.
$2,499 and $18,995 would buy a lot of hookers.
The Roadster Sport is pretty cool
I don't understand guys. Why would chaining these things together give us higher performance than having a bunch of ordinary CPU's?
Lots of transistors.
Because it runs on the power of AWESOME!!11
CPUs are good at branching (decision making) not so much at math. GPUs are the exact opposite, extremely good at math, bad at branching.
It's like asking why you wouldn't want to use humans to crank out cars from scratch or why robots with the advanced AI of today can't compose symphonies. I'm sure if you got enough of one or the other, you could do it, but it would cost a shit-ton.
Get my point?
dagamer43, That's also the exact reason why these will never encroach upon true Super Computing Space. Most of the Top 500 Super Computers are used in areas where branching is a very common thing. Really the areas that these will excel in are DSP based. Signal Processing, etc. These are very bad for state-space searches, which is what Scientific Super Computing is really for.
@Shyam That's not exactly how it works. CUDA is a model for heterogeneous computing, which basically means that information is passed back and forth between the CPU layer and the GPU layer. Teslas can handle minor branching without it being much of a problem (the penalty is that it has to evaluate both branches regardless). While there is some significant branching that takes place in scientific computing, there is usually a large amount of computing between those points. So what we do is mate 1 / 2 CPU cores to each GPU, let the CPU decide what operations need to be performed, then send the data and instruction to the GPU. GPU does its work, then returns results to the CPU to decide what to do next. And yes, this is *way* faster than standalone CPU's.
@Shyam, lots of varied operations are done with scientific supercomputers, including extremely large matrix inversion and repetitive vector calculations. Vector processors excel at these operations. It all depends on what you're trying to accomplish... using the right tool is important, but there is definitely a place for this technology in high performance computing.
This is kind of off topic, but it makes me sad to think of all the trolls I've seen on Engadget as of late. The comment section has turned into a cesspool of the internet, 4-Chan 2.0 if you will. Here's hoping that comenters think before they type (no one in particular here, BTW).
Way to stay on topic!
Thanks.
You know, calling all of engadget's comm enters trolls in and of itself is trolling.
What really gets to me though is people who call spammers and flamers trolls, which they are not. There is a difference.
*thinking*
There's truth to this, but we don't really have an appropriate forum or venue to talk about these issues either.
Maybe an Engadget forum would be a good idea.
Did they name it after Nikola Tesla, the greatest inventor of the world?
No, the name comes from what hides under the hood - Tera-electronvolt Energy Superconducting Linear Accelerator...
[/sarcasm]
You mean Mr. Nokia Tesla.
Nice to see these finally making an appearance. I can say from some experience that using the TESLAs in our supercomputers at school has lowered the price a bit (About $10K). Not to mention made them somewhat more scalable.
I know dell and HP started to sell computers with them already installed and configured for about $16K, which isn't bad at all. Compared to what it used to be.
Delicious wood screws.
I'm going to calculate so many prime numbers with this thing, like 1, 3, 7, and 8.
I can't believe I actually laughed for like 3 seconds after reading that. Well played.
They haven't been "tossing around" CUDA, genius, there's actually apps that make good use of it to cut down on H.264 rendering (Adobe Premiere Plugins, BadaBoom H.264 converter, etc.)
Hahahaha. Maybe this doesn't matter to a supercomputer, but $2500 and only one video out port?
You're right, it doesn't. Downranked.
At that price, I think people will continue to be inclined to purchase a stack of PS3's and a metal rack.
But can it run Crysis?
I'm surprised no one has said that yet...
In all seriousness though, this is pretty cool. I can't wait till CUDA is a common thing in applications and I can get a decent boost from my graphics card. Well, once I buy one, that is.
Or maybe these things would bake a badass renderfarm.
In all seriousness, I do wonder how well/badly this sort of GPU handles gaming. Obviously, your typical or even high-end consumer is not really the target audience for this, but would its raw power outweigh those considerations?
Pointless, I suppose, but I'm still curious.
@Prolorn It would be good, but nothing amazing. The current-get Tesla board (the C1060) is pretty much a GTX295 with a lower clock speed and 4GB of onboard RAM.
Throw couple of those in your rig and MD5 and lesser SHA1 become quite readable. Even my aged 8800GTX can crunch 300 mil. MD5s / second without breaking a sweat... Well, at least we know that NSA would be ordering the the first batch of those.
A Tesla card with video output? Say it isn't so!
(Seriously, what is up with that output?)
Actually it's a very nice feature. Have you tried putting a cluster full of tesla cards together? If your motherboard doesn't have onboard video, the only way to debug these things are via ssh, serial connection (also motherboard dependent), or by putting in a graphics card. I like the single DVI port idea, so I can get a KVM to the machine.
@SmittyMcSmith
Psst smithy, I think he was having a jab at the recession.