bit-tech.net

Nvidia CEO reveals Fermi architecture

Nvidia CEO reveals Fermi architecture

GTC 2009: Nvidia CEO Jen-Hsun Huang revealed the company's next generation GPU architecture, codenamed Fermi.

Nvidia's CEO and President Jen-Hsun Huang has revealed Nvidia's next-generation GPU architecture, which goes by the codename Fermi.

It represents the first major architectural change since the company announced its G80 graphics processor in 2006.

Fermi packs three billion (yes, you read that right) transistors and features 512 CUDA cores with suppor for IEEE 754-2008 floating point precision. It has native support for C++, which complements existing support for C, Fortran, Java, Python, OpenCL and DirectCompute.

There's also a new cache hierarchy, known as Nvidia Parallel DataCache, and an updated Nvidia GigaThread Engine with support for concurrent kernel execution. Nvidia has also introduced Nexus, which it claims is the world's first integrated heterogeneous computing application development environment within Microsoft Visual Studio.

High performance computing users haven't been left out dry either, as double precision throughput has increased by a factor of eight compared to the previous generation. Nvidia has also added support for ECC memory, which is a critical requirement for datacentres and supercomputers looking to deploy GPUs on a large scale.

There's quite a lot to take in here, and we've not got the full details yet - we'll be doing a much fuller analysis as soon as we know more. One key thing to consider is graphics performance, but we're quite confident here as Huang said that "we will make real-time ray tracing a reality this year."

We have to wonder where some of these changes leave AMD in the compute space because, as we mentioned in our Radeon HD 5870 Architecture Analysis, AMD's GPU computing strategy isn't as strong as it could be compared to Nvidia's (before Fermi) and now the holes become more apparent.

Discuss in the forums.

29 Comments

Discuss in the forums Reply
Isitari 30th September 2009, 23:42 Quote
hhmmm Games are definitely not going to need THIS much power. The ATI 5870 is already well overpowered for anything now, that's why AMD are pushing using Many multiple displays as this is the only thing that actually pushes the hardware!

Isitari.
glaeken 30th September 2009, 23:53 Quote
There really isn't any thing as too much power. This much power is great for other areas besides games: offline computing, scientific visualization, etc.
Threefiguremini 1st October 2009, 00:00 Quote
How do people ever buy hardware for their pc?? Every time I come close to buying a part there's a new one just over the horizon!
RotoSequence 1st October 2009, 00:00 Quote
It seems strange to position yourself in the market of co-processors that do double duty as graphics cards. It's expensive and there are hundreds of millions of transistors that will inevitably be useless to the majority of their market...

Do they really think that selling millions of co-processors to people will keep their company afloat against more dedicated solutions? It seems a bit hard to imagine that this will work out for them. Oh well, its an interesting thing to see. :)
Tim S 1st October 2009, 00:04 Quote
One of the demos they've shown is speeding up early tumour detection for breast cancer... going from approx 9 hours to less than 30 mins. Nvidia's moving beyond just graphics and has been for some time... it looks like this time they've not forgotten about graphics like they seemed to with GT200.
tejas 1st October 2009, 00:07 Quote
This is looking good particular for CUDA programming close to metal. This is definitely geared more towards compute than gaming which is not necessarily a bad thing.

Folding anyone???
Diosjenin 1st October 2009, 02:55 Quote
Anand put up a relatively in-depth look at the new architecture earlier today, including why they're emphasizing compute. Interesting stuff.
Saivert 1st October 2009, 03:28 Quote
lol, it's a long time ago since NVIDIA was purely a graphics company. They are into CPUs (ARM), GPUs, CUDA, and all that now.
This way they can have a foot in more markets and that will make them stay afloat.

ATI can ride the cheap graphics bandwagon as long as they want of course (we like cheap graphics).
Elton 1st October 2009, 04:30 Quote
The sad thing is that they aren't focusing on Graphics which is pretty much the focus of the GPU industry.
l3v1ck 1st October 2009, 04:46 Quote
Ever since the 3Dfx Voodoo, GPU development has always been about making games run faster with better graphics.
Reading this, this is the first time I've felt that's not true any more.
I think Nvidia can see future profits coming from large business rather than PC gaming.
Elton 1st October 2009, 05:02 Quote
Quote:
Originally Posted by l3v1ck
Ever since the 3Dfx Voodoo, GPU development has always been about making games run faster with better graphics.
Reading this, this is the first time I've felt that's not true any more.
I think Nvidia can see future profits coming from large business rather than PC gaming.

Makes sense, although I want GPUs to be like the 6800GT ---> 7800GTX. Or the 7800GTX ---> 8800GTX.

I mean the evolution was incredible. Now it's still a bit dissapointing. I mean the 8800GTS 640/320MB was almost 2x the performance of the previous generation, and remember this was a midrange card.
DiegoAAC 1st October 2009, 05:53 Quote
I'm guessing that Fermi will be 10-20% faster than Evergreen, but nVidia will make ~30% less money per gpu than AMD (2x10^9 vs. 3x10^9 transistors at 40 nm TSMC). Also that the capability to execute code written in c++ is through LLVM.
Itbay 1st October 2009, 06:43 Quote
When you will decide to buy a device the next day you will find another better than that...just confusing..:?
thehippoz 1st October 2009, 06:46 Quote
huang said he's going to start offering suki suki with each nvidia card.. this reminds me so much of when the g80 was released- everyone was like holy cow waiting for ati's response

ati and that guy at the inquirer were hyping up ati's phantom card for the longest time.. it released and everyone who was waiting like myself was like wtf- it was just a bunch chicken ****

I'm not holding my breath.. real time ray tracing.. smoke up my ass
Adnoctum 1st October 2009, 11:05 Quote
I always take Charlie's nVidia predictions with a grain of salt, but I'm starting to think he may be right about a Q1 2010 launch for "GT300" this time.
The rumours floating around is that the GT300 is floating around developers and so on. To quote Johnny Rotten: "Horses, Horses, Horses**t!"
Do we all seriously think that Jen-Hsun "Whoop-Ass" Huang would present Fermi to the public and NOT show it in action? Especially when he is attempting to dull the Evergreen launch?
And no, I don't count an image of it supposedly ray-tracing a Bugatti (are we to take nVidia's word on this?) and holding a black card in his hand that does nothing.
The problem I have at the moment is that I don't trust nVidia even when they may be telling me the truth.
Too much FUD, not enough action.
Adnoctum 1st October 2009, 11:30 Quote
Have you guys checked out the comments of the Anandtech Fermi article? There's the biggest Internet troll raging insanely I've ever seen called SiliconDoc. He's hilarious, and so incredibly sad. Is this the same SiliconDoc banned here? I can't help but notice a similarity.

Article comments: http://www.anandtech.com/video/showdoc.aspx?i=3651#comments
Bit-Tech profile: http://forums.bit-tech.net/member.php?u=41666
mclean007 1st October 2009, 11:36 Quote
Quote:
Originally Posted by DiegoAAC
I'm guessing that Fermi will be 10-20% faster than Evergreen, but nVidia will make ~30% less money per gpu than AMD (2x10^9 vs. 3x10^9 transistors at 40 nm TSMC). Also that the capability to execute code written in c++ is through LLVM.
I don't think you can put much stock in those figures. First, a very large portion of the unit cost of a GPU, especially in its early production phases, must relate to amortisation of research and development costs.

Secondly, I would expect the cost to produce a 3bn transistor chip will be significantly more than 50% more than the cost to produce a 2bn transistor chip - if yields were perfect, that might be the case, but remember that even one dud transistor can spoil a chip entirely, and the likelihood of a 3bn transistor chip having at least one dud is substantially greater than for a 2bn transistor chip.

Finally, even if your assumptions on cost of production were right, that is simply cost of production. It has to be assumed that nVidia will charge a premium for its GPU if it is faster than AMD's offering, so revenue per unit will be higher. This higher price would be expected to have a knock on effect on volumes.

There are too many unknowns in this for you or me to make any kind of estimate as to which architecture will be more profitable.
tejas 2nd October 2009, 00:49 Quote
Screw crap x86 CPU garbage from AMD and Intel.

Fermi and Cypress are beautiful works of art and mark the start of the GPU age. Kudos to ATI and Nvidia for keeping well ahead of Moores Law and giving developers and programmers new toys to play with!
general22 2nd October 2009, 14:19 Quote
Well they seemed to have slightly more than doubled the compute power so for games so it should end up reasonably faster than the 5870. Too bad its still a little while off. Hopefully NVIDIA aren't being completely stupid and abandoning performance in 3D Graphics since that is basically 99% of their revenue.

I also read somewhere that the demo fermi card shown was just a mock up so I'm guessing a release will not happen this year.
John_T 2nd October 2009, 19:30 Quote
Quote:
Originally Posted by Threefiguremini
How do people ever buy hardware for their pc?? Every time I come close to buying a part there's a new one just over the horizon!

Very true, but it's a bit like swimming in a cold pool: If you want to swim, then at some point you just have to pinch your nose and jump in...
thehippoz 2nd October 2009, 20:15 Quote
this guy makes a good observation at the bottom of the page

http://www.semiaccurate.com/2009/10/01/nvidia-fakes-fermi-boards-gtc/

the solder pins don't line up underneath the 8 pin bottom of the board XD

&edit- doh kroten beat me to it in the other thread =]
LordPyrinc 3rd October 2009, 03:58 Quote
Give me a video card that isn't the size of a netbook!!!!

My PC is only 2 years old now, but I had a hard time finding a decent card to fit in it when my 8800 GT card died. Seriously.. the graphics cards coming out these days are about as long as the motherboard and usually take up two slots. Makes it a b**ch if you have any connections on the motherboard that fall under the massive footprint of the newer video cards.

I admit that buying an off-the-shelf pc limits me somewhat, but really??? Do these cards need to be this f*n big? Can't they at least make them shorter? Why the hell do the cards need to be 10-12 inches long?
Elton 3rd October 2009, 05:59 Quote
Quote:
Originally Posted by LordPyrinc
Give me a video card that isn't the size of a netbook!!!!

My PC is only 2 years old now, but I had a hard time finding a decent card to fit in it when my 8800 GT card died. Seriously.. the graphics cards coming out these days are about as long as the motherboard and usually take up two slots. Makes it a b**ch if you have any connections on the motherboard that fall under the massive footprint of the newer video cards.

I admit that buying an off-the-shelf pc limits me somewhat, but really??? Do these cards need to be this f*n big? Can't they at least make them shorter? Why the hell do the cards need to be 10-12 inches long?

I'll say yes if you can find a way to do it. It's not that they even waste space on it, it's that cards nowadays require so much room because there's so many components in them.
SchizoFrog 4th October 2009, 18:03 Quote
Quote:
Originally Posted by LordPyrinc
Give me a video card that isn't the size of a netbook!!!!

My PC is only 2 years old now, but I had a hard time finding a decent card to fit in it when my 8800 GT card died. Seriously.. the graphics cards coming out these days are about as long as the motherboard and usually take up two slots. Makes it a b**ch if you have any connections on the motherboard that fall under the massive footprint of the newer video cards.

I admit that buying an off-the-shelf pc limits me somewhat, but really??? Do these cards need to be this f*n big? Can't they at least make them shorter? Why the hell do the cards need to be 10-12 inches long?

Blame the Motherboard manufacturer's for not thinking about high end graphics when they design board layouts. they could have componants where the GPU sits and move the connectors to another place on the board.
feathers 12th October 2009, 16:13 Quote
"Isitari 30th September 2009, 22:42
hhmmm Games are definitely not going to need THIS much power. " - Games expand to fill the capability of the hardware. I do remember my uncle telling me in 1983 that games would never need more than 64k memory. Do you really think game technology will stop here and never become more complex?

This kind of graphics technology means the potential for better game AI as well as more realistic imagery. We haven't even started on ray-traced games yet and even Fermi is probably a little too weak for that.
feathers 12th October 2009, 16:15 Quote
"RotoSequence 30th September 2009, 23:00
It seems strange to position yourself in the market of co-processors that do double duty as graphics cards. It's expensive and there are hundreds of millions of transistors that will inevitably be useless to the majority of their market...

Do they really think that selling millions of co-processors to people will keep their company afloat against more dedicated solutions? It seems a bit hard to imagine that this will work out for them. Oh well, its an interesting thing to see. :)" - Of course they think it. Unlike you they have a clearer vision of what is and isn't feasible. Their solutions are many times cheaper than the dedicated solutions you talk about. Very inexpensive and scalable. No contest. Nvidia wins.
thecrownles 3rd November 2009, 18:56 Quote
...512 CUDA cores. I'd be surprised if they didn't find a way to error-test based on groups of cores and their shared cache and disable the error cores, selling lower-end cards with bad blocks of disabled cores to increase their yield. There must be a way to do this using different board firmwares for different working core configurations.
Dreamcometru 12th November 2009, 18:37 Quote
whats real time ray tracing mean?
Log in

You are not logged in, please login with your forum account below. If you don't already have an account please register to start contributing.



Discuss in the forums