nVidia GT300's Fermi architecture unveiled: 512 cores, up to 6GB GDDR5

[IntelUser2000]

Originally posted by: lopri
So Larrabee = $200? Interesting.

This is interesting if its true. How do they manage to have it profitable with a >500mm2 die and at that price(mind you, that's the final board cost, not the chip). Probably its not meant to be, hmm.

The successor will have to be >>>performance and <<<die size to be profitable. Maybe its not meant to be actually. Just a "strategic" investment.

 

[bunnyfubbles]

As awesome as this sounds, I'm not sure I like what I'm seeing from nVidia in the grand scheme of things, I really haven't been that impressed with their mid range parts since the G92 first came out. Although there definitely looks like there is some room to be pleasantly surprised here.

Originally posted by: dguy6789
Only 384 bit memory? What happened to the 512 bit + GDDR5?

overkill?

GDDR5 will already provide more than double the bandwidth of GDDR3, no need to inflate the costs of the card even more if the bandwidth isn't needed.

Its already painfully obvious this is absolutely a monster of a chip, of which if they followed the same path as the GTX280, it would mean they would want to charge $600+ for it. 512bit memory interface would mean more even complex card designs which drives up the cost.

Originally posted by: happy medium

Originally posted by: OCguy
That sounds more credible than anything Charlie has ever wrote, but I just want benchmarks at this point.

1.5GB memory standard? Sweet!

Guess we can speculate the lower end cards (gt350) will have 768mb memory? Mabe the 3gb card is the gtx395?

Lower end cards might not have 384bit memory interfaces. If this is anything like the 8800GTX and GTS days we'll see 320bit (1280MB) and probably 256bit (1024MB), as GDDR5 really does provide a ton of bandwidth, there wouldn't be much need to keep the more elaborate interfaces. I think it would be foolish to have anything other than the true budget parts have anything less than 1GB of memory.

 

[zebrax2]

Would Intel and AMD patent severely hamper their move to general computing? Or does it not apply because it isn't x86?

 

[dookulooku]

Originally posted by: Kakkoii

Cypress= Press Codename
RV870= Manufacturing codename
HD 5870= Retail name when put together as a video card.

Fermi= Press Codename
GF180= Manufacturing codename
GTX 380= Retail name when put together as a video card.


People were throwing around the term "GT300", because the manufacturing codename of the 2XX series was "GT200",

HD 5870 / GTX 380 = Marketing name.

You ought to read Anandtech's articles more carefully.

Evergreen is the name of the architecture for ATI's DX11 chips. Cypress is the codename of the chip. I don't think the name RV870 is used anymore.

Tesla is the name of the architecture for NVIDIA's DX10 chips. GT200 is the name of the chip; the "T" stands for Tesla. So Tesla is both an architecture and a product line for HPC solutions.

Fermi is the name of the new architecture, so the chips could be called GFxxx yet the product for the HPC market could still be called Tesla.

 

[Griswold]

Does nobody wonder how much of the transistor budget is left for ROPs and TMUs after all the GPGPU enhancements that went into this design? I refuse to believe there will be a massive design difference (other than ECC functionality maybe) between the tesla and geforce products.

Nvidia is clearly moving its focus away from gaming over to GPGPU - and thats a good plan. PC gaming is a shrinking market whereas GPGPU/HPC is where the beef is at in the future.

 

[Keysplayr]

Originally posted by: dookulooku

Originally posted by: Kakkoii

Cypress= Press Codename
RV870= Manufacturing codename
HD 5870= Retail name when put together as a video card.

Fermi= Press Codename
GF180= Manufacturing codename
GTX 380= Retail name when put together as a video card.


People were throwing around the term "GT300", because the manufacturing codename of the 2XX series was "GT200",

HD 5870 / GTX 380 = Marketing name.

You ought to read Anandtech's articles more carefully.

Evergreen is the name of the architecture for ATI's DX11 chips. Cypress is the codename of the chip. I don't think the name RV870 is used anymore.

Tesla is the name of the architecture for NVIDIA's DX10 chips. GT200 is the name of the chip; the "T" stands for Tesla. So Tesla is both an architecture and a product line for HPC solutions.

Fermi is the name of the new architecture, so the chips could be called GFxxx yet the product for the HPC market could still be called Tesla.

I thought "Evergreen" was the term used for the 6 display port cards? And standard models are referred to as "Cypress or CypressXT"

SAPPHIRE 100281SR Radeon HD 5870 (Cypress XT) 1GB 256-bit GDDR5 PCI Express 2.0 x16
SAPPHIRE Radeon HD 5870 (Cypress XT) 100281SR Video Card

 

[Keysplayr]

Originally posted by: Griswold
Does nobody wonder how much of the transistor budget is left for ROPs and TMUs after all the GPGPU enhancements that went into this design? I refuse to believe there will be a massive design difference (other than ECC functionality maybe) between the tesla and geforce products.

Nvidia is clearly moving its focus away from gaming over to GPGPU - and thats a good plan. PC gaming is a shrinking market whereas GPGPU/HPC is where the beef is at in the future.

Somebody mentioned the possibility, or rather the concept, of a single Fermi being the heart of a new console. Imagine if all that was needed beyond this card, is a board to plug it into that has an I/O chip, disk controller, USB interface and LAN chip. Concept, remember. I'm sure it's not at all as simple, but it does make one think of the possibilities.
I think Nvidia is closer to a real "Fusion" or Larrabee concept, and it kinda crept up on us.

 

[cusideabelincoln]

Originally posted by: Keysplayr

Originally posted by: dookulooku

Originally posted by: Kakkoii

Cypress= Press Codename
RV870= Manufacturing codename
HD 5870= Retail name when put together as a video card.

Fermi= Press Codename
GF180= Manufacturing codename
GTX 380= Retail name when put together as a video card.


People were throwing around the term "GT300", because the manufacturing codename of the 2XX series was "GT200",

HD 5870 / GTX 380 = Marketing name.

You ought to read Anandtech's articles more carefully.

Evergreen is the name of the architecture for ATI's DX11 chips. Cypress is the codename of the chip. I don't think the name RV870 is used anymore.

Tesla is the name of the architecture for NVIDIA's DX10 chips. GT200 is the name of the chip; the "T" stands for Tesla. So Tesla is both an architecture and a product line for HPC solutions.

Fermi is the name of the new architecture, so the chips could be called GFxxx yet the product for the HPC market could still be called Tesla.

I thought "Evergreen" was the term used for the 6 display port cards? And standard models are referred to as "Cypress or CypressXT"

SAPPHIRE 100281SR Radeon HD 5870 (Cypress XT) 1GB 256-bit GDDR5 PCI Express 2.0 x16
SAPPHIRE Radeon HD 5870 (Cypress XT) 100281SR Video Card

Evergreen refers to all the presumably 5000 series cards - Cypress, Juniper, Redwood, etc.

 

[Keysplayr]

Originally posted by: cusideabelincoln

Originally posted by: Keysplayr

Originally posted by: dookulooku

Originally posted by: Kakkoii

Cypress= Press Codename
RV870= Manufacturing codename
HD 5870= Retail name when put together as a video card.

Fermi= Press Codename
GF180= Manufacturing codename
GTX 380= Retail name when put together as a video card.


People were throwing around the term "GT300", because the manufacturing codename of the 2XX series was "GT200",

HD 5870 / GTX 380 = Marketing name.

You ought to read Anandtech's articles more carefully.

Evergreen is the name of the architecture for ATI's DX11 chips. Cypress is the codename of the chip. I don't think the name RV870 is used anymore.

Tesla is the name of the architecture for NVIDIA's DX10 chips. GT200 is the name of the chip; the "T" stands for Tesla. So Tesla is both an architecture and a product line for HPC solutions.

Fermi is the name of the new architecture, so the chips could be called GFxxx yet the product for the HPC market could still be called Tesla.

I thought "Evergreen" was the term used for the 6 display port cards? And standard models are referred to as "Cypress or CypressXT"

SAPPHIRE 100281SR Radeon HD 5870 (Cypress XT) 1GB 256-bit GDDR5 PCI Express 2.0 x16
SAPPHIRE Radeon HD 5870 (Cypress XT) 100281SR Video Card

Evergreen refers to all the presumably 5000 series cards - Cypress, Juniper, Redwood, etc.

Ok, so the whole gambit.

 

[MarcVenice]

Mighty interesting. I hope for Nvidia that there is a big HPC-market they can sell their gpgpu's too.

First though: Considering this puppy will most likely be 500mm2 or bigger (i've heard 576mm2), it will consume quite a bit of power, 200W tdp or more most likely. Putting two of them together is probably not doable. Meaning Nvidia will have to use a cutdown version, like with the GTX295, to make a dualgpu card. This gpu should be faster then a HD 5870, so a slower version will probably still be faster or at least be equal to a HD 5870. That's going to put Nvidia and AMD really close in terms of the performance crown. Also, with the gpu being this big yields will be lower then AMD's, making the gpu even more expensive. Pricewise Nvidia needs a good deal with TSMC...

Second thought: When saying that I hope the HPC-market is big enough for Nvidia to sell a lot of gpgpu's, I say that because with a gpu this big, they'll have a hard time adressing the mainstream market. And the mainstream gpu-market is where all the money is, besides a possible HPC-market and professional market. (ATI's sales in the professional market only account for a few percent of their total videocard sales, but the margins are so big, it earns them 10 percent or so of their total reveneu, so there is definately money to be had there).

 

[BenSkywalker]

Fermi not only looks like it a great Tesla chip, but it pretty much looks like MS or Sony could design a console around a single Fermi chip and some RAM and call it a day.

Getting a bit ahead of the curve atm

While GF100 is a rather large step in that general direction, a couple of major architectural issues are going to prevent that in no uncertain terms. The biggest one, no matter the computational capabilities of the chip, is the amount of cache available. With a maximum of 48KB per 'core' and 768KB L2 for 16 'cores' general purpose code is still going to roll over and die on this chip. To compound this problem each 'core' has 32 functional units, the entire architecture of this chip is very much tailor suited to large scale FP calculations, and not very well suited to large batches of simple operations. If you absolutely must pick one or the other for a general purpose CPU, you go with the latter which is where the GF100 would fall down, hard.

That said, this chip paired with even the existing(and now very cheap) Cell or POWER would make for a fairly decent console. Of course, give this another 1.5-2 generations to bake and toss on a little eDRAM for kicks and then you could be in business, and I'm sure that is something nVidia is keeping in their minds as we close in on the timeframe when the big two are looking for a new supplier for the next gen.

except larrabee is x86. should work right out of the box on some stuff.

Larrabee is in order and requires heavy vectorization to get decent performance. You aren't going to get very much to run decently on it without some work.

I think Nvidia is closer to a real "Fusion" or Larrabee concept, and it kinda crept up on us.

Speak for yourself Keys, I've been calling it for a while now

First though: Considering this puppy will most likely be 500mm2 or bigger (i've heard 576mm2)

I believe Anand said it was 476mm, smaller then the GT200(he explicitly stated that).

it will consume quite a bit of power, 200W tdp or more most likely.

The board they were showing had a single 8 pin, I don't pay enough attention to power draws to worry about that too much, just pointing out an observation(I would assume if it was <200W it would require multiple plugs?).

Second thought: When saying that I hope the HPC-market is big enough for Nvidia to sell a lot of gpgpu's, I say that because with a gpu this big, they'll have a hard time adressing the mainstream market.

It's smaller then the GT200, and they took ~66% of the market with that generation.

 

[Zstream]

Originally posted by: Keysplayr

Originally posted by: Griswold
Does nobody wonder how much of the transistor budget is left for ROPs and TMUs after all the GPGPU enhancements that went into this design? I refuse to believe there will be a massive design difference (other than ECC functionality maybe) between the tesla and geforce products.

Nvidia is clearly moving its focus away from gaming over to GPGPU - and thats a good plan. PC gaming is a shrinking market whereas GPGPU/HPC is where the beef is at in the future.

Somebody mentioned the possibility, or rather the concept, of a single Fermi being the heart of a new console. Imagine if all that was needed beyond this card, is a board to plug it into that has an I/O chip, disk controller, USB interface and LAN chip. Concept, remember. I'm sure it's not at all as simple, but it does make one think of the possibilities.
I think Nvidia is closer to a real "Fusion" or Larrabee concept, and it kinda crept up on us.

Emm no, you can not run an OS off of Cuda or OpenCL.

 

[BenSkywalker]

Emm no, you can not run an OS off of Cuda or OpenCL.

Not an existing OS, you could do it with the capabilities of this chip(although not in a graphics card configuration) but it wouldn't be very good at it.

 

[Kingbee13]

Initially I had the same idea of a console on a chip as others have considered, but I quickly dismissed the idea, current game engines can't even saturate 8 threads on a chip in the ~3ghz range, how would they do on a much lower clocked chip that supports 24,576 threads?

 

[WelshBloke]

While its certainly a good move for Nvidia I think its a shame they are moving away from traditional video cards into the GPGPU market.

It would be nice if maybe they had two products, a 'pure' videocard for people like me who don't really care about the CUDA stuff and a GPGPU version for people who do.

Originally posted by: Nvidia
For sixteen years, NVIDIA has dedicated itself to building the world?s fastest graphics
processors. While G80 was a pioneering architecture in GPU computing, and GT200 a major
refinement, their designs were nevertheless deeply rooted in the world of graphics. The Fermi
architecture represents a new direction for NVIDIA. Far from being merely the successor to
GT200, Fermi is the outcome of a radical rethinking of the role, purpose, and capability of the
GPU.

 

[MarcVenice]

First though: Considering this puppy will most likely be 500mm2 or bigger (i've heard 576mm2)

I believe Anand said it was 476mm, smaller then the GT200(he explicitly stated that).

Anand extrapolated this number by assuming 40% more transistors then Cypress will inexplicably result in a 40% bigger diesize. Cypress = 334mm2, times 1.4 = 467,7mm2. Now Nvidia sais it's OVER 3 bill transistors. It's a good a guess as any, but 576mm2 is what I heard from my own sources. Only time will tell the exact number.

it will consume quite a bit of power, 200W tdp or more most likely.

The board they were showing had a single 8 pin, I don't pay enough attention to power draws to worry about that too much, just pointing out an observation(I would assume if it was <200W it would require multiple plugs?).

The board they were showing doesn't even have to be the high-end gpu. And, an 8pin can supply 150W (in fact, an 8pin can deliver up to 240W, or so i've been told) Anyways, assuming 150W, +75W from the pci-e slot, means 225W. TDP could still be 200W. (oh and I'm not sure, but don't pci-e 2.0 slots deliver 150W?) In that case we're talking 300W total.

-edit: New information, although the foto only shows a 8pin, there's also a 6pin on the card, somewhere. This would once again allow for a minimum of 300W (8pin =150W, 6pin =75W, pci-e slot =75W)

Second thought: When saying that I hope the HPC-market is big enough for Nvidia to sell a lot of gpgpu's, I say that because with a gpu this big, they'll have a hard time adressing the mainstream market.

It's smaller then the GT200, and they took ~66% of the market with that generation.

You 'think' it's smaller, because of a educated guess anand made. And they did NOT take 66% of the market with gt200, they took 66% of the total market, where g92b probably took MOST of the mainstream market for Nvidia, where as gt200 only accounted for the upper mainstream and high-end market.

 

[Keysplayr]

Originally posted by: Zstream

Originally posted by: Keysplayr

Originally posted by: Griswold
Does nobody wonder how much of the transistor budget is left for ROPs and TMUs after all the GPGPU enhancements that went into this design? I refuse to believe there will be a massive design difference (other than ECC functionality maybe) between the tesla and geforce products.

Nvidia is clearly moving its focus away from gaming over to GPGPU - and thats a good plan. PC gaming is a shrinking market whereas GPGPU/HPC is where the beef is at in the future.

Somebody mentioned the possibility, or rather the concept, of a single Fermi being the heart of a new console. Imagine if all that was needed beyond this card, is a board to plug it into that has an I/O chip, disk controller, USB interface and LAN chip. Concept, remember. I'm sure it's not at all as simple, but it does make one think of the possibilities.
I think Nvidia is closer to a real "Fusion" or Larrabee concept, and it kinda crept up on us.

Emm no, you can not run an OS off of Cuda or OpenCL.

Are you 100% sure? Because I,m not 100% sure that it couldn't. Doesn't have to be Windows or any other common OS.

 

[Kuzi]

Originally posted by: BenSkywalker

Emm no, you can not run an OS off of Cuda or OpenCL.

Not an existing OS, you could do it with the capabilities of this chip(although not in a graphics card configuration) but it wouldn't be very good at it.

Fermi would probably run ok as a CPU on a decade old console like a Dreamcast for example

 

[lavaheadache]

So basically what I got out of the article is that you can hot wire a hard drive to the video card and install an OS... lol.


While I'm excited about this sure to be beastly video card' I'm rather annoyed with intent of this leak.

One word came to mind, desperation. Nvidia might have well stated something like this. "please folks, remain calm with your credit cards in your wallets. Don't waste any of your available credit on an inferior hd5870 because you will need all probably need to ask for a one time credit increase at the time of fermi release."

I imagine by the time this behemoth (and I mean that in regards to size, performance, features..etc, etc) gets released, an X2 version and possibly some kind of hd5890 will be ready to go much like the x1900 came out to put some whoop ass on the 7800 GTX 512.

As a neutral consumer/enthusiast, I sure an getting awfully tired of nvidia and they're constant strategies/antics they play on us lab mice.

 

[Janooo]

Originally posted by: BenSkywalker
...

I think Nvidia is closer to a real "Fusion" or Larrabee concept, and it kinda crept up on us.

Speak for yourself Keys, I've been calling it for a while now

...

It's smaller then the GT200, and they took ~66% of the market with that generation.

The idea of Fusion and Larrabee is x86 + GPU. NV does not have x86. They would have to create their own CPU or buy/license something.

66% is not correct. They sold less GT200 than ATI RV7xx.

 

[Janooo]

Originally posted by: BenSkywalker

Emm no, you can not run an OS off of Cuda or OpenCL.

Not an existing OS, you could do it with the capabilities of this chip(although not in a graphics card configuration) but it wouldn't be very good at it.

QFT. Fermi would stink if it had to run an OS.

 

[Janooo]

Originally posted by: Keysplayr

Originally posted by: Zstream
...
Emm no, you can not run an OS off of Cuda or OpenCL.

Are you 100% sure? Because I,m not 100% sure that it couldn't. Doesn't have to be Windows or any other common OS.

100%
OS needs I/O subsystem, file subsystem, interrupts, ...
Are they present in Cuda or OpenCL? I don't believe they have such libraries.

 

[BFG10K]

Originally posted by: Griswold

PC gaming is a shrinking market whereas GPGPU/HPC is where the beef is at in the future.

No it isn?t; PC gaming is actually growing but most of the time revenue isn?t tracked properly, namely that of digital distribution and MMO subscriptions.

 

[MarcVenice]

Some must read articles for those who'd like to have a clue about Fermi:

http://www.realworldtech.com/p...icleID=RWT093009110932
http://www.techreport.com/articles.x/17670

And definately check out this thread: http://forum.beyond3d.com/showthread.php?t=55289&page=1

 

[Compddd]

Originally posted by: BFG10K

Originally posted by: Griswold

PC gaming is a shrinking market whereas GPGPU/HPC is where the beef is at in the future.

No it isn?t; PC gaming is actually growing but most of the time revenue isn?t tracked properly, namely that of digital distribution and MMO subscriptions.

This!

Also, if this doesn't work out like Nvidia hopes, and it backfires on them. Are they screwed?