Nvidia's Future GTX 580 Graphics Card Gets Pictured (Rumours)

[Idontcare]

Is there a particular limit to how large a GPU AMD or nV can make at TSMC?

If we are restricting the scope of discussion to that of a monolithic IC then yes there is...the limit is the full-field exposure window of the lithography tool.

And since there are many different litho tools employed in the manufacture of a single IC, it is specifically the minimum field of exposure of all the litho tools.

Generally the maximum exposure field is 29mm x 29mm (~800mm^2). They aren't always square and they aren't always so large.

Not too mention that for practical purposes you never want to use the entire exposure field.

But Nvidia certainly has room to grow their GF100 chip if they wanted.

Now if we wanted to consider variations on the theme of bolted together GPUs that aren't monolithic then the sky (i.e. the budget) is the limit, GPU's can be larger than a wafer after MCM assembly if you were willing to pay for it (the technology exists).

 

[dangerman1337]

If we are restricting the scope of discussion to that of a monolithic IC then yes there is...the limit is the full-field exposure window of the lithography tool.

And since there are many different litho tools employed in the manufacture of a single IC, it is specifically the minimum field of exposure of all the litho tools.

Generally the maximum exposure field is 29mm x 29mm (~800mm^2). They aren't always square and they aren't always so large.

Not too mention that for practical purposes you never want to use the entire exposure field.

But Nvidia certainly has room to grow their GF100 chip if they wanted.

Now if we wanted to consider variations on the theme of bolted together GPUs that aren't monolithic then the sky (i.e. the budget) is the limit, GPU's can be larger than a wafer after MCM assembly if you were willing to pay for it (the technology exists).

Even if it could fit, i think the results will be not liked by a lot though .

 

[cody_horner]

Why do all the posts on this site turn into some kind of stupid argument? lol I've not read one thread on here where someone hasn't had a pickle up their ass.

That being said, I'm curious then if this card might be better for gaming whereas the current GTX480 might be better for GPGPU operations?

 

[Nemesis 1]

ooooooo

 

[RussianSensation]

No way,
TMUs are too large to double them.

GF100 SM:
16 CUDA cores (#1)
16 CUDA cores (#2)
16 Load/Store Units
16 Interpolation SFUs (not on NVIDIA's diagrams)
4 Special Function SFUs
4 Texture Units

GF104 SM:
16 CUDA cores (#1)
16 CUDA cores (#2)
16 CUDA cores (#3)
16 Load/Store Units
16 Interpolation SFUs (not on NVIDIA's diagrams)
8 Special Function SFUs
8 Texture Units

They increased the complexity of each SM in terms of TMUs, without increasing die space (we have seen them do the same with 8800GTS 320/640 (G80) --> 8800GT (G92). NV has shown that they can increase TMU performance, while reducing die size with both G92 and GF104 chips.

"For GF104, NVIDIA has doubled up on the number of texture units. A “full” GF104 has the same number of texture units at GF100 (64) in half as many SMs. "

The reduction in GF100-->GF104 space came from removal of GPGPU components/reducing # of shaders/reducing FP64/double precision and tessellation engines. So you cannot just assume that doubling of TMUs/performance results in doubling of the size.

How can that be? Are you harping on him for using common sense? Am I missing something? If you take Fermi, which is already a huge die, and double the number of TMU's, you are adding even more to an already huge die.

It doesn't stand to reason that increasing TMU performance will result in increased die size over the 529mm2 GF100 die size. Unless of course you were talking about increasing of GF104 die size?

GTX460 vs. GTX470.
332mm2 die size vs. 529mm2 die size.

GTX460 has 11% higher texture fill-rate and 25% less SPs than a GTX470 despite a 37% smaller die.

So any assumptions about NV being unable to increase texture fill-rate performance aren't supported by any "logic" or "common sense".

Didn't the GF100 had a lot of features on it that were HPC based rather than gaming whihc took up die size? I think Jen hsung talked about this at GTC 2010 conference about the fermi design on why so big and hot etc.

Yes . Moreso, it's impossible to compare GF104 SPs to GF100 SPs.

In a best-case scenario GF104 can utilize 4 of 7 execution units, while GF100 could only utilize 2 of 6 execution units. The ability to extract ILP from a warp results in GF104’s compute abilities performing like a 384 CUDA core part some of the time, and like a 256 CUDA core part at other times, despite only having 336 SPs. It will be less consistent, but on average faster than a pure 256 CUDA core part would be.

GF100 --> GF104:
1. NV removed ECC.
2. For GF104, NVIDIA removed FP64 from 2 of the 3 blocks of CUDA cores. The effective execution rate is 1/12th FP32 performance vs. 1/8th for GF100.

They increased TMUs, reduced other things, and ended up with a proportionally smaller die.

GF104 --> full chip (384 cores) vs. GF100 --> full chip (512 cores). 75% of SPs with 2x as many TMUs per SP in a die size that's 63% of the original size is disproportionally smaller.

Alternatively, 336 SP GTX460 vs. 480 SP GTX480 has 70% of SPs, and 90% of Texture Fill-Rate performance of GTX480 in 63% of the die size.

 

[chewietobbacca]

They increased the complexity of each SM in terms of TMUs, without increasing die space (we have seen them do the same with 8800GTS 320/640 (G80) --> 8800GT (G92). NV has shown that they can increase TMU performance, while reducing die size with both G92 and GF104 chips.

Flaw in the analogy: G92 was a shrink to 65nm from the G80s, so of course die size went down.

"For GF104, NVIDIA has doubled up on the number of texture units. A “full” GF104 has the same number of texture units at GF100 (64) in half as many SMs. "

The reduction in GF100-->GF104 space came from removal of GPGPU components/reducing # of shaders/reducing FP64/double precision and tessellation engines. So you cannot just assume that doubling of TMUs/performance results in doubling of the size.

Interestingly enough, over at B3d, more than a few posters claimed GF100 initially had 128 TMUs but 64 were cut out to improve yields/whatever. The GF104 having the same amount despite being less overall might correspond quite well to that fact. Its possible GF110/GTX580 has fixed whatever issue was there and is back to 128, but if some leaked synthetics are true, then they might be at 64 again.

 

[RussianSensation]

NV knows GF100 is lacking texture fill-rate (GTX470 has less than GTX285). AMD's advantage really starts to show at 2560x1600 with HD5850 often > GTX470, HD6870 easily > GTX460 and GTX480's lead almost diminishing over HD5870. More or less, increasing texture fill-rate is a must for GTX580 imo. Otherwise, HD6970 would crush it with ease at higher resolutions.

 

[Arkadrel]

@RussianSensation

big cards like 6970 and 580 only really make sense for people playing high resolutions... so yeah nvidia will have to mainly improve their texture fill-rate.

Amd has ample texture fillrate... id like to see them work on their tessellation <.< to be able to go toe to toe with a 580 at x32+ factors. Rumor has it they improved the tessellation in the 69xx cards by a good bit so maybe we ll see that.

 

[Matrices]

@RussianSensation

big cards like 6970 and 580 only really make sense for people playing high resolutions... so yeah nvidia will have to mainly improve their texture fill-rate.

Amd has ample texture fillrate... id like to see them work on their tessellation <.< to be able to go toe to toe with a 580 at x32+ factors. Rumor has it they improved the tessellation in the 69xx cards by a good bit so maybe we ll see that.

Why even bother with tessellation? The inclusion of DX11 features by either company has proven a waste of die resources. Developers don't care and few games - very few games, actually - incorporate DX11 features in any meaningful way.

A feature like MLAA is more welcome - along with Eyefinity, it was a smart recognition that game development has irrevocably shifted to consoles and game developers don't give a shit about PC hardware, so the best thing to do is to just optimize what we can get out of the console-oriented development processes.

 

[Sickamore]

THE DESPERATION AT Nvidia has reached silly proportions, with the paper launch of the GTX580 pulled in from 'Cayman Day' to 'Investor Conference Call Day', both holidays in November. I wonder how many reviewers will overlook the fact that there won't be cards this year?

We told you about the GTX580 a few weeks ago and now a few more details have come out. The short story is that the chip behind the GTX580, the GF110, is nothing more than a bug fixed GF100. If you think this should be called the GF100b, you are right. If you think performance will underwelm, you are right too.

Sources deep within Nvidia tell SemiAccurate that the clock speeds targeted by GTX580 are 750/1500, or the exact same frequencies that Nvidia was targeting before the GF100/GTX480 missed. Where have we heard that number before? Oh yeah, we have been saying that that was the original clock target for GF100 for over a year, and Nvidia has been denying it for just as long. They denied many other problems too, but that's not important now that we have an imminent launch.

It is pretty obvious that GF100b/GF110 was supposed to be a shrink to 32nm, but when TSMC canceled that node, Nvidia had to jump into action on Plan B. That meant there was time to do a base layer respin on GF100 to fix some of the bugs.

A fully working GF100, a first for Nvidia's desktop Fermi/GF1x0 line, would have 512 shaders and get about a 7% speed increase over GTX480. Another 6% comes from the clock speed, a generous assessment would say bug fixes add a little more. All told, 20% net speed increase over GTX480 isn't out of the question, nor is a pretty decent lowering of power.

The chip is still going to be obscenely huge, and, well, it won't be enough to beat the competition. Heck, it won't be enough to beat the current champ, AMD's HD5970/Hemlock, much less the upcoming Cayman and Antilles parts. Is is going to be a slaughter.

That is why Nvidia is pulling the 580 launch in, if you launch before the competition, effectively there is no competition. That means the good press from tame sites and analysts will go unchallenged, and that is what Nvidia desperately wants. Correction, it's what NVDA needs.

Why do they need it? Because they are about to have a horrid quarter, and need good news. The past few weeks have been a whirlwind of backdoor deals and heavy discounts at quarters end, including some previously sacred cows. AMD is kicking them right between the margins, something you will see when Q4 numbers come out. The best way to deflect questions aimed at management is to show Wall Street something shiny.

With that in mind, we are told GTX580's 'launch' will be pulled in to November 8, a few days before the Q3 financial conference call. It is now meant as a spoiler for uncomfortable analyst questions aimed at Dear Leader, not at AMD's parts. The problem is that in either case, you won't be able to buy parts until late January, best case, at the earliest. I wonder if any analysts will ask about that in an SEC governed venue?

In the end, GTX580 will be non-existent in 2010, for you, the customer, and is aimed purely at those who are too dumb or cowered to question Nvidia's word. The chip is a bug fixed GTX480/GF100, renamed on the internal and consumer side to make it seem like something it isn't, ala the endless G92 stream or the GT3xx line. We hope the wait for these cards is less than the wait between paper launch and real cards was for Fermi.S|A


http://www.semiaccurate.com/2010/10/30/nvidia-paper-launch-gtx580-week/

 

[notty22]

How come Charlie never comments on AMD's financial reports ?
But seems obsessed with Nvidia's, lol

 

[tviceman]

If there is no availability this year, then that will be ridiculous. However, to say a 20% improvement in speed coupled with a 20% decrease in power consumption is a failure within the same node is ridiculous. Even if this is the chip GF100 was supposed to end up being, and keep in mind Charlie always low balls performance estimates from Nvidia chips, then it's still going to be a competitive chip.

It will be interesting to see if this all pans out to be true.

 

[AtenRa]

I thought TMUs are going up from 64 to 128?

It doesn't stand to reason that increasing TMU performance will result in increased die size over the 529mm2 GF100 die size

Fact 1. Doubling the number of TMUs will raise the die size because you double the transistor count a TMU is comprised off.

Fact 2. We can increase the Texture performance without increasing the number of the TMUs. That can be achieved by redesigning the TMU and make it more efficient or increase the operational frequency.



If you think that GTX580 will be a GF100 derivative, then doubling the TMUs from 64 to 128 will increase the die size.

If you think that GTX580 will be a GF104 derivative, then doubling the TMUs from 64 to 128 by doubling the whole GF104 is understandable.

I was talking for the first, because I believe GTX580 will be a GF100 derivative with 512 Cuda Cores, 64 TMUs and 48 ROPs.

 

[Arkadrel]

The 580 *might* end up bigger than 529mm^2.

at Semiaccurate a few think the 6970s will probably 380mm^2+ size.
With amd haveing almost twice as much performance/mm^2... well Im guessing over 529mm^2 if it follows a GF100 derivative.

big chips = expensive sales prices (because of higher cost makeing)
(expection being when someone sells something at a loss)

I dont doubt that nvidia can make a chip thats 20&#37; faster than a 480 though, I doubt it ll be a small chip => expensive prices on it.

 

[tannat]

Fact 1. Doubling the number of TMUs will raise the die size because you double the transistor count a TMU is comprised off.

Fact 2. We can increase the Texture performance without increasing the number of the TMUs. That can be achieved by redesigning the TMU and make it more efficient or increase the operational frequency.



If you think that GTX580 will be a GF100 derivative, then doubling the TMUs from 64 to 128 will increase the die size.

If you think that GTX580 will be a GF104 derivative, then doubling the TMUs from 64 to 128 by doubling the whole GF104 is understandable.

I was talking for the first, because I believe GTX580 will be a GF100 derivative with 512 Cuda Cores, 64 TMUs and 48 ROPs.

There were strong rumours half the TMUs were disabled on GTX480. 128 TMUs falls within this possibility and may still be possible even if GTX580 would end up being a respin of GF100.

I find this unlikely though, since, to my knowledge, nvidia have never changed the chip designation (gf100 to Gf110) with only a respin. Even a respin with dieshrink would only give it a b at the end. Like G200 to G200b.

 

[RussianSensation]

I was talking for the first, because I believe GTX580 will be a GF100 derivative with 512 Cuda Cores, 64 TMUs and 48 ROPs.

Ya, if it's still based on GF100, then I doubt they'll double it to 128, but why would they still base it off GF100? Why not just take GF104 chip and build from there instead? I guess it's just cheaper for them to have built up inventory of full GF100 512mb / 64 TMU chips. Now they will just increase clock speed due to better transistor maturity. That would be pretty disappointing if the 580 is just a GF100 chip. Perhaps, they will improve the efficiency of each Cuda Core and TMU?

NV removed 300 million transistors related to HPC when moving from GF100 to GF104. They can simply remove ECC, reduce FP64/double precision performance, and focus on creating a gaming chip this time around as well. At this point it's all just rumours.

 

[SickBeast]

Nvidia's track record of late has been to produce as big of a bad-a** GPU as they physically can on a die. Die size and efficiency are not their goals with high end GPUs. They just want it to be the fastest thing out there, everything else be damned.

Their midrange derivatives are a different story, but at the high end, I definitely see the logic in the speculations many people here are making regarding the GTX580's die size.

 

[AtenRa]

Well, I believe they will continue with GF100 (Fermi) style architecture for their high End chip. The reason is, they want the High End chip to be a GPGPU and Gaming (Tessellation) chip at the same time.

Yes I believe the efficiency will be better in GTX580 and with more shaders + raised frequency + better Efficiency they will be able to get 20%+ performance vs GTX480 and if they will be able to decrease power usage too it will be a nice and a welcome surprise.

 

[RussianSensation]

They just want it to be the fastest thing out there, everything else be damned.

hehe if NV spent 8 months revising bugs on GF100 to only release another 250W+ card that need 90&#37; fan speed to keep it ~ 85*C, while being only 20% faster at $499, then their management truly lost it in 2010. That's how they plan to gain market share when there are 0 demanding games coming out from now until Crysis 2 at the end of March 2011?

Niether NV nor AMD have anything worth buying > $260. So wouldn't it be more logical to release full GF104 @ $250-$350 cards first? I guess for the 10 people who buy $500 videocards every 6 months, it will be another card to aim for 60fps on Crysis 1920x1080 8AA bench... :whiste: 20% extra performance over 480 is not going to make a dent in Metro 2033.

 

[Scali]

Well, I believe they will continue with GF100 (Fermi) style architecture for their high End chip. The reason is, they want the High End chip to be a GPGPU and Gaming (Tessellation) chip at the same time.

Well, theoretically they could trim down the architecture a bit, like they did with GF104. After all, they already have the GF100 to serve the GPGPU/HPC market. They could continue selling that alongside the GTX580 which would be more concentrated on gaming.

 

[SickBeast]

hehe if NV spent 8 months revising bugs on GF100 to only release another 250W+ card that need 90% fan speed to keep it ~ 85*C, while being only 20% faster at $499, then their management truly lost it in 2010. That's how they plan to gain market share when there are 0 demanding games coming out from now until Crysis 2 at the end of March 2011?

Niether NV nor AMD have anything worth buying > $260. So wouldn't it be most logical to release full GF104 @ $250-$350 cards first? I guess for the 10 people who buy $500 videocards, it will be another card to aim for 60fps on Crysis 1920x1080 8AA bench... :whiste: 20% extra performance over 480 is not going to make a dent in Metro 2033.

NV is pigeon-holed into making big bad-a** GPUs at this point. They can't make chipsets for Intel (for the most part), AMD now does their own chipsets in-house, and you have the newer CPUs shipping with integrated GPUs that NV once sold in large quantities.

NV looks like they are betting on GPGPU tech catching on, and they're making the fastest GPUs that they can because it's the only part of the market that the "Fusion" type chips will not devour over time.

 

[SHAQ]

How come Charlie never comments on AMD's financial reports ?
But seems obsessed with Nvidia's, lol

Exactly.

http://forums.anandtech.com/showthread.php?t=2115596

 

[AtenRa]

Well, theoretically they could trim down the architecture a bit, like they did with GF104. After all, they already have the GF100 to serve the GPGPU/HPC market. They could continue selling that alongside the GTX580 which would be more concentrated on gaming.

Well yes they can do that, but you know what ?? for me GF104 was a step backwards and not forward.

Yes it is better for today’s games, it has better power usage and it is cooler than its bigger brother but, it sacrificed a lot of its tessellation capabilities in order to get there and it became a lot more like AMDs offerings.

Don’t get me wrong, GF104 is a fine chip, but technology must go forward and we need the right hardware with new innovations to push the new generation of technology. I would prefer a new GF100 derivative with an updated polymorph engine to raise tessellation performance vs GTX480 than a new chip that it will be faster in today’s games.

Just my 2 cents

 

[Scali]

Yes it is better for today’s games, it has better power usage and it is cooler than its bigger brother but, it sacrificed a lot of its tessellation capabilities in order to get there and it became a lot more like AMDs offerings.

Not sure about that, really.
In my opinion, they just scaled down the tessellation capabilities to match the rest of the chip's performance (not much point in having a full GTX480 tessellator on there when the rest of the chip cannot render the triangles that fast anyway).
The GTX460's tessellation performance is still above that of AMD's fastest offering in that field: the 6870.
As long as GF110 has tessellation scaled to its performance as well, I'd be fine with that.

 

[Arkadrel]

Scali that depends on which tessellation benchmark and which factors your running. At the moderate levels in ungine it shows like this:

Moderate tessellation used: 460 EVGA FTW (OCed) beats 6870 (stock) by 10&#37;.
Extreme tessellation used: 460 EVGA FTW (OCed) beats 6870 (stock) by 20%.

Now this a unfair compairison because the 460 EVGA FTW is heavly overclocked vs stock speeds 6870. The 460 oced still isnt beating the 6870 by gigantic amounts. I remember seeing some reviews where they used 460 (stock) vs 6870 (stock) and the 6870s where beating the 460s in Ungine benchmarks at both moderate and extreme levels.


*** notice that the super overclocked 460 evga ftw is also beating the 470 in tessellation. Is that a fair way to compair it against a 6870 at stock speeds?