NVIDIA had told us it would be
accelerating its CUDA program to try and get an advantage over its competitors as
OpenCL brings general-purpose GPU computing to the mainstream, and it looks like that effort's paying off -- the company just announced its new Fermi CUDA architecture, which will also serve as the foundation of its next-gen GeForce and Quadro products. The new features are all pretty technical -- the world's first true cache hierarchy in a GPU, anyone? -- but the big takeaway is that CUDA and OpenCl should run even faster on this new silicon, and that's never a bad thing. Hit up the read links for the nitty-gritty, if that's what gets you going.
Read - NVIDIA Fermi site
Read - Hot Hardware analysis
Read - PC Perspective analysis
posted Oct 1st 2009 2:09AM
first!
SONY, Please make this a PS4 graphics chip
judging by how poorly the ps3 is doing, i dont know if we'll even see a ps4
To say the PS3 is doing poorly is incorrect, sir.
It may not be selling like the PS2 but it is doing fine.
I think Sony probably learned their lesson and will be on top of the console game when the next gen rolls around.
The post should mention that this is a paper launch and we won't be seeing any usable Fermi chips until at least Q1 of 2010.
This thing is going to be expensive to make and will not scale up very well.
You must work for AMD.
Integrated porn is the next step for a 3rd Gen processor.
I wanna be a millionaire to buy that latest stuff monthly!;)
Wow that's the weirdest looking chip ever, it almost looks like art or something.
good tech is always art
Wow, all the above comments were just terrible. Did any of you read the article? I guess 457R4L might have... Now, I'll actually read the article.
...
Hopefully the new architecture brings some kind of power savings to new Geforce models, as the cards have become so ridiculously large and power hungry. Just replaced my old 7900gs with a 9400gt, and I couldn't believe how big the thing was.
Um, you've probably noticed this in benchmarks already.. but the 9400GT probably is a step down from a 7900GS in performance, although it's got a larger featureset.
i would like to see this applied to a hybrid cpu
After watching that video, all I want to know is if this will finally make ray tracing possible at decent fps and resolution.
don't know if this means anything but their optix engine does look promising for raytracing especially in CUDA.
http://www.nvidia.com/object/optix.html
the commenting demographic clearly shows who engadget is catering towards now.
hurrrrrr
I see trouble coming to AMD/ATI's way... CUDA's gonna crush them hard in their face. Just like their slogon: The way it's meant to be played. RIP ATI, we'll miss you...
I see fanboys making utterly moronic comments...it's already happening, oh God!
So now there are ATI/nVidia fanboy wars?
Gosh--kind of pathetic...
If I had to pick a side, I'd pick to be a 3DFX fanboi.
@rickjamess04:
What do you mean, "now"?
hmm, I thought about building a new desktop with an ATI graphics card, because they aren't that expensive, faster for less money and the latest cards have a very low idle power consumption, but all those cards don't have CUDA and I would really like to have a CUDA GPU for Photoshop and other future software.
Does ATI have anything similar, else the decision is easy: A mid NVidia with CUDA.
I just have the feeling that NVidia cards just offer more features than the ATI cards.
Pretty much.
Comparing raw numbers, the Fermi and the 5xxx series from AMD look to be about equal in terms of theoretical peak numbers. But when you look at what NVIDIA is saying about Fermi's memory architecture, they've left AMD in the dust on this one. Multilevel coherent caches, high performance global atomics? Tripled shared memory space?
The problem with ATI is basically they're just a bunch of smart chip designers, who got bought by AMD, who are also just a bunch of smart chip designers. They've always been good at coming up with fast, efficient chip designs in a pinch. But they've never really understood that there is more to the business than pixel fill rate. Yes, you get a device which has fantastic raw numbers, but in almost every other way the product is inferior. Frankly, I always get the sense that they'd be perfectly content if they could exist without customers, pouring over specifications and perfecting their chip designs. Admirable in a way, but not terribly useful in the real world.
An example of this: ATI's drivers traditionally stank to high heaven, so even when they had the faster chip (which happens quite a bit less often than ATI fans would like to admit), they were spanked by NVIDIA's better driver stack. It took them years to understand the importance of a software stack as devices moved from pure hardware to a mixture of hardware+software. Move on a few years into the future: When fully programmable GPUs came around, they were again caught with their pants down slavishly perfecting their implementation of the previous generation's technology. And their missing of the boat on GPGPU is nothing short of epic. Even while they were touring the circuit bragging about their CUDA killer a couple years ago, they couldn't ship sample units to save their lives, let alone produce a useful software stack to go with it.
The 5xxx vs the GT300 looks to be more of the same. The 5xxx is a technically brilliant solution to last year's problem. The current big problem with GPUs isn't that they're not fast enough, it's that they're only fast on a narrow class of problems. We don't need a GPU which is faster on that same set of problems, we need a GPU which is as fast on a broader set of problems (I'll take faster if I can get that too). From the announced feature set, it sounds like the GT300 is going to go a long way towards solving this year's problem. The 5xxx can be 10 times faster than the GT300 for all I care, if I can solve the same problem on the GT300 with an O(n) algorithm when it requires an O(n^2) algorithm on a 5xxx because of architectural limitations.
jepzilla, I agree with you. But.
"The 5xxx is a technically brilliant solution to last year's problem."
Yes, and there is nothing wrong for company to sell a card which greatly improve performance of my actual game library.
nVidia's feature set is very very good, but in the end they cost too much, updated too often and most software (even games) rarely exploit the feature set to the fullest.
Worst part of that is the by the time software gets updates to use all the features of your -now affordable- nVidia card, they craftly release new generation of cards, marking your card essentially obsolete. They did it before with Riva/TNT to GeForce transition. They did again with 7xxx -> 8xxx transition. And they are apparently set to do that again with 2xx -> 3xx transition.
I'd like to love nVidia, but as long as they keep concentrating on higher end graphic cards - with low/middle end being a mere afterthought - I'm going to stick with my decision to switch to ATI (I went gf7800 to 4850). Prices for 58xx already started falling (actually in Europe the cards started below MSRP) and I can't justify spending 400-500€ (where usually new nVidia cards start) on something what only few exotic pieces of software would fully exploit.
"We don't need a GPU which is faster on that same set of problems, we need a GPU which is as fast on a broader set of problems."
And that seem to be solution to no problem. Or - an expensive solution to a niche problem. Video encoding - only widely used example I can come up with - is going to be faster, yet it is a niche task.
nVidia simply tries to compete with Intel. They they are going to loose (unless of course they come up with GPU which can also replace CPU). Kudos to AMD/ATI which instead try to make better cards for games with have.
"and I would really like to have a CUDA GPU for Photoshop and other future software."
I have (soon to be replaced) 3870s, and they've support OpenGL since January - acceleration for rendering images is already
available with CS4.
Except it doesn't really matter how much better Fermi is compared to the 5000 series since it's not a hard launch with actual products, just a paper launch of a future architecture that's coming.
I mean it took them what, a full year to get Ion to market from "launch"?
Anyways for now it's nice that they have something but until actual hardware is in stores I'm not going to hold my breath.
Hmmm, openCL or cuda. I'm cheering for the one that runs on both vendors and not just nvidia's.
I am curious to see openCL run on ATI cards to see just how they compare to cuda
More PR smoke & mirrors from Nvidia, you'll notice not a mention of Direct Compute under Win7/Vista either which means they've probably done the numbers and found AMD's DX11 5000 series kicks Fermi's ass. AMD/ATI has always been better at executing the full Direct-X profile than Nvidia who only concentrate on core areas and leave performance lacking in others.
Read the tech info !!! fermi supports DC, dose so very well. Actually all three DC OpenCL and CUDA have the same basic principals, CUDA is the most advanced since it has been around for much longer, and has only one body making the decisions for its feature set. But fermi will kick ass on DC.
For how much NVIDIA trash talks about Intel, they sure are moving closer and closer to turning the GPU into a full on CPU. NVIDIA scoffed at Larabee when Intel announced it, but seeing as how Fermi is cache coherent, how is it really that much different from Larabee?
> ... seeing as how Fermi is cache coherent ...
... I welcome CUDA programmers to the hell of multi-threaded programming - which surely would ensue.
OpenCL lol
cuda gone --;
Now... If they would like to lower the power consumption and heat.
It's nice to see technology breakthrough, but sustainability is very important as well...
I'd also like to point out that this is mostly just 2 GTX280 chips put together, expect high power consumption, high heat, and high price
You didn't read the article did you?
What a Schmuck. XD
I really don't think this is any kind of deal breaker for ATI at all, the few sites that I have read about this paper release today all pretty much were iffy on this chip actually being faster or more powerful than the latest ATI release.
The only thing that Nvidia has going for it is CUDA and maybe Physx, but as an Nvidia SLi user myself I think Physx is really only utilized in a handful of games properly and for the consumer side CUDA support is only used properly in a handful of titles as well. I've tried several CUDA accelerated video encoding programs and have always gone back to my old programs due to lower quality output and glitches in the videos that I don't experience with other programs. On the other hand Nvidia has never supported the Direct X fully and their video card drivers quality is all over the map and it's a dice roll as to whether updated drivers will help or harm your system.
I think that OpenCL and Direct Compute will be bigger items going forward than CUDA and to a lesser extent DirectX 11 support so I really don't see anything here thats going to adversely effect ATI at this point.
I've been doing a tick-tock pattern with my motherboard/cpu and SLi cards for the the last few years and I am on my 5th SLi setup at this point so I'll probably stay with Nvidia this next generation also but I do think ATI will be just as effective going forward with the emerging API's as Nvidia in the GPGPU arena.
As somebody in the industry, here's my perspective on OpenCL/DirectCompute vs CUDA. Nobody is interested in DirectCompute. I'm sure it'll find some usage in game engines, but beyond that it's practically off the radar. OpenCL is a more interesting proposition, and there is broad interest in OpenCL. But at the same time, the view of OpenCL from the trenches isn't quite as rosy because most developers expect that kernels are going to have to be custom written for specific target architectures anyway, for performance reasons. Worse, OpenCL is not as featureful as CUDA when it comes to GPGPU.
On the business side the idea of being able to target multiple hardware platforms with one API is very appealing, since it avoids NVIDIA lockin. But today NVIDIA is the only viable platform, but AMD's offerings in the field have always been a bit academic; maybe that'll change over the next couple of years, but it's an unknown today. Worse, CUDA is already worming its way into some very conservative fields - it's not uncommon to see brand new CUDA code working alongside 20+ year old fortran code, to give you an idea how receptive some of these fields are to rewriting code - and it has the more mature skillset and toolkit.
I think ultimately though, the deciding factor will be the OpenCL standards process and OpenCL 2.0. It's enlightening to consider the analogue with OpenGL & Direct3D. OpenGL was the popular choice, supported by popular vendors, Direct3D a lockin platform by one fairly unpopular company that was arguably technically inferior (certainly in its earliest revisions). Yet Direct3D dominates over OpenGL in most areas, because Microsoft was able to develop the API fast enough to keep pace with rapidly improving hardware capabilities. OpenGL stagnated in committee. Today, OpenCL is a more limited platform than CUDA. Given the popularity and interest I've seen it could certainly deplace CUDA, but if the Khronos group treats it the same way it treated OpenGL it will die on the vine.
Actually cuda supports opencl and direct compute and is based on the c language so these will also be in macs, i just hope these things are smaller than the last generation because they are getting ridiculously huge, hot, power hungry
I hope there will be some realtime raytracing coming out soon. That would just be awesome.
www.raytracing.ch
www.openrt.de
www.realstorm.com
and so on...
I think in 2-3 years we see the first cool ray tracing game...
We all know this thing will be faster than the new Radeon by a bit in some applications, be more expensive, use more power, larger, hotter but will be the king. Then ATI will come out with the 5890... be on top again then Nvdia will come out with the GTX385 and again be on top and again be more expensive. This is the pace ATI and Nvidia have set up so for the foreseeable future ATI will be almost as fast as Nvidia cards but be cheaper too. I prefer Nvidia :)
Over its competitor, not competitors.
Dear Mr Engadget,
"NVidia's next generation of GPUs will be faster than the previous one, so it's good for OpenCL"
What the hell happend to you ? Did a TUAW virus eat your brain or something ?
Please take care and heal quickly, and come back with lots of useful and informative articles as we know you can write.
viruses do not exist in TUAW... OS X remember?
So, a 5850 for my gaming PC and a GT300 for modeling oil fields and super novas.
Cool.
This looks like a good chip with some innovative ideas but in the real world for someone looking to buy a new graphics card the 58xx series is available now and its benchmarks give you a clear idea what you're paying for.
When this eventually reaches the retail market and we get to see what it's price / performance ratio is, only then will it be worth comparing. Until then it's just another engineering sample . . . .
Who the hell wouldn't wait a month or two for Nvidias offering? It can be a 400 dollar decision and one that most people have to live years with. You would be an idiot not to wait for the 300 reviews first.
And I mean that so you can make a decision based on all the options out there. The 300 might not be as expensive since ATI set the price bar first, so Nvidia has no choice but to match that depending on the performance of the card.
Personally, I'm waiting for the 300 reviews because I'm not a fanboy and I'm not an idiot. Then ill make my purchase for the best card out there I can afford.
Except it's not going to be a month or two; more like 3-5 months.
Right now, my new P55 i7 is limping along on a $40 (post-rebate) 9600 GSO. I did this knowing I wanted to wait for the DX11 cards. ATI's are here, nVidia's aren't; combine that with nVidia's chips that destroy themselves and the recent PhysX disabling debacle, it's not hard to make the ATI choice.
Nvidia's problem is in the main stream market. ATI is already offering DirectX11 cards in the higher end and mid range market. By years end ATI will have its DirectX offerings in the main stream and value market with $50-150 cards and that's also where the big money is. Customers that buy those offerings don't care about OpenCL/DirectCompute/CUDA all they see is a higher DirectX number in the support list and higher benchmark numbers they will choose ATI over Nvidia. There could be some very tough times head for Nvidia.
Not interested.
I'm not going to support CUDA or PhysX with my money, i'll only support open technologies like OpenCL, Open Physics etc.
*buys a 5850*
CUDA supports OpenCL.
*returns your 5850 for you*
But who supports CUDA? Well there's Nvidia and.... Nvidia. Oh, right.
I won't support CUDA unless it goes open so Intel, ATI etc. can support CUDA programs.
*cancels return*
@Reuben
A month or two? $400? Did you even bother to read the article or the visit the Nvidia links?
They've only announced the *architecture*. There's nothing to indicate a retail product based on this technology will be available for sale in 2 months or that it will priced around the $400 mark. That's just pure speculation on your part.
Ati/AMD have a product on sale now that will deliver a quantifiable performance level for a known price.
You could just as easily wait 6 months for Fermi architecture to appear in a retail product and then find it's priced nearer the $600 mark. My point is, until an actual product is released anything else is just guess work.
If you need/want the fastest DX11 graphics card available now, the 58xx series is a good choice. If you want to wait for the next great innovation in graphics technology (and Fermi looks pretty interesting) then by all means do so but to suggest anyone who buys now is an idiot makes your comments discourteous at best.
the playstation 3 was supposed to have real time raytracing capabilities, from a research paper I read in 2005. Doesn't that use an NVIDIA chip? Well we all know how that worked out.
FOUR DEE GAMING! C'mon, don't lump nvidia into Sony's empty promises. PS3's marketing revolved around deceipt, showing pre-rendered video in place of actual game play. Talk about smoke and mirrors.
I doubt Sony is going to make the same mistake in the future. Maybe for the next gen they won't be so terrified of the competition, pre-release.
It can do it. You just need 3 of them:
http://www.youtube.com/watch?v=oLte5f34ya8
And I'm pretty sure that every single resources of the cpu is used for rendering meaning no games possible....
how about you start by getting a job instead of sitting on engadget. then you can buy whatever you want because gpus are only a few hundred dollars.
from what i see it looks like Nvidia is trying to bypass using Intel CPUs in the future nvidia is laying down the Base for a ture GPGPU Like Intel's Larrabee.
Looks like Apple's new technological innovations on the parallel computing front are beginning to sprout with the fantastic management skills of the Khronos group.