GTX 480 with 512 shaders 'soon'?

[qubit]

Digitimes has a rather extensive review covering a wide range of subjects, but the first two questions and the answers are really the most interesting ones. The first question relates to the missing shaders, i.e., no 512 shader GPUs, and although Drew Henry isn’t willing to discuss any specifics, he does go on record saying that “[Nvidia has] a chance to launch a graphics chip with 512 cores in the future.” This doesn’t really say much, but at least it indicates that Nvidia is working on getting out some new products that should hopefully perform better than the GTX480.

I wonder how soon is soon?

SemiAccurate (don't worry, it's not Charlie D reporting. )

EDIT: The merest mention of Charlie D seems to get tongues wagging, but please don't discuss him here, as it's not the topic for this thread. Please keep the discussion about an enhanced Fermi architecture.

 

[Benetanegia]

I wonder how soon is soon?

SemiAccurate (don't worry, it's not Charlie D reporting. )

You can already know it's not Charlie by looking at the first word on the quote above... "Digitimes has a...". A source instead of a link to another Charlie's article... yeah, not Charlie.

Regarding the 512 SP version, I think that everybody knew that it would come eventually, but I think it's going to be a replacement like the 216SP GTX260 instead of an addition* and that's something I don't like too much. It's not that bad and both ATi and Nvidia have done that plenty on the past (often times it was the clocks that changed), but it's not ideal either.

*That way 480, 470 and 460 will have 64 SPs difference from each other.

 

[DrPepper]

Hmm more heat and power draw. This will go down well with the ATi crowd. I think nvidia should have just released it with 512 shaders to begin with.

 

[kid41212003]

It's not that they didn't, but they couldn't instead.

 

[DrPepper]

It's not that they didn't, but they couldn't instead.

Some of those die's can do 512 shaders in the same sense that GTX460 and 470 could if they weren't disabled. They've done it to increase yields and increase product availability but I still haven't found any for sale in the UK.

 

[W1zzard]

they are collecting dies now i guess .. at an estimated rate of 5 chips per month, so maybe next year we can have a limited edition card

 

[TVman]

sorry but i dont care about interviews with PR people if it was nvidias one of the top engineers i would take them seriosly but a PR "GUY" these people are payed to spur out endless diarrhea.

 

[sandiegoborn32]

Hmmm....Nvidia better come up with a better cooling system before they put more shader processors in.

More heat = less hardware life...

 

[DrPepper]

Hmmm....Nvidia better come up with a better cooling system before they put more shader processors in.

More heat = less hardware life...

Chances are the PCB in any gpu will fail before heat destroys anything. I don't think many people have the issue of cards failing due to heat these days but I could be wrong.

 

[sandiegoborn32]

Chances are the PCB in any gpu will fail before heat destroys anything. I don't think many people have the issue of cards failing due to heat these days but I could be wrong.

Yeah, in my opinion, heat is a computer hardware's worst enemy, and that's why I try to keep my hardware as cool as possible.

I might even need to get water cooling soon, if I keep improving my system with new hardware and more overclocking...

 

[overclocking101]

everyone talks crap about charlie but facts are facts and 90% of what he said was correct about fermi.

 

[qubit]

everyone talks crap about charlie but facts are facts and 90% of what he said was correct about fermi.

Absolutely, +1. He seriously knows what he's talking about and he's had some blindingly good articles (all rants, of course). However, he does sometimes just talk crap - the 3D Vision stuff comes to mind here.

Way back on The Inquirer, he said it was gonna be rubbish. Then he saw it at a trade show when it was new and revised his opinion and said it looked quite good (still at The Inquirer). This seemed fair enough and quite unbiased at that point.

However, he then screwed up recently at SemiAccurate by saying that 3D Vision is crap and pointing to his original article, ignoring the second one - no way has he forgotten his actual experience with it and his second article. He can and should be pulled up and flamed over this.

Benetanegia said that he's seen other stuff like this from Charlie too and I believe him.

 

[lism]

they are collecting dies now i guess .. at an estimated rate of 5 chips per month, so maybe next year we can have a limited edition card

Lol.

So a chip with the full 512SP's enabled would cause over 300+Watts of power usage?

I think nvidia is re-adopting the same chip in a smaller process... This way it'll should draw less power and able to hit up some more Flops.

 

[newtekie1]

Personally, I think people focus far too much on the GTX480 and the GTX470 gets pushed to the side. When in reality, I think the GTX470 is the winner GPU. It has decent power and heat aspects, not great like the HD5800 series, but decent. It performs well still, and is in-expensive compared to the GTX480 and HD5870.

As for a 512 SP part coming, it is probably almost certain, but I'm not worried about it. It will be even more expensive and put out even more heat than a GTX480, so I don't want it.

I'm really looking forward to the revamped next generation GPUs more than a 512 SP Fermi.

Benetanegia said that he's seen other stuff like this from Charlie too and I believe him.

Basically every GPU generation he rants about how the new nVidia GPU will never see a part with all the SPs enabled. He did it with G80(and of course the 8800GTX/Ultra had all the SPs enabled), he did it with G92(and of course the 9800GTX/8800GTS512 had all the SPs enabled), and he did it with G200(and of course the GTX280 had all the SPs enabled). He has once again done it with Fermi, and we will have to see what comes in the future, but I'm not going to be surprised if we see a GTX490(or whatever) with all 512 SPs enabled.

But really this thread isn't about Charlie, so lets try to keep discussion about him out of it and focus on the topic.

 

[Benetanegia]

everyone talks crap about charlie but facts are facts and 90% of what he said was correct about fermi.

Let's forget about the fact that the things he said were going to be bad and are bad are not as bad as he said they would be and just concentrate on the things he was wring about:

- He said that Fermi would not have ROPs and obviously has.
- He said it would not have TMUs and obviously has.
- He said it would not have a tesselator and it has 16.
- He said it would not have DX11 and make it through software emulation. DX11 runs much better than DX9/10 so much for being emulated isn't it?

There you have 4 things he was absolutely wrong. Now you tell me 36 things in which he was right please? The only thing he was completely right on is that it was going to be late and have low yields and that's something anyone could anticipate. The thing about yields is obvious, but are not as bad as he says. Hell, his own claims about yields never made any sense at all. I mean he has said many times that Nvidia ordered 9000 risk waffers for Fermi and that yields were below 10% and on the same page he said that Nvidia would have less than 8000 cards out. Let's see 9000 waffers and less than 8000 cards, that's less than 1 die per waffer and every Fermi card would cost like $5000 to make, which is obviously false. Also his recent reports about yields and all are false. At below 20% Nvidia would be losing more than $200 per card and they have not been in red numbers on their quarter revenues, they've done better than in past year in fact. They are not selling GT200 cards either, so from where is Nvidia getting the money in order to cover all those gigantic loses? 2+2=4 they are not having such loses.

 

[qubit]

Personally, I think people focus far too much on the GTX480 and the GTX470 gets pushed to the side. When in reality, I think the GTX470 is the winner GPU. It has decent power and heat aspects, not great like the HD5800 series, but decent. It performs well still, and is in-expensive compared to the GTX480 and HD5870.

As for a 512 SP part coming, it is probably almost certain, but I'm not worried about it. It will be even more expensive and put out even more heat than a GTX480, so I don't want it.

I'm really looking forward to the revamped next generation GPUs more than a 512 SP Fermi.



Basically every GPU generation he rants about how the new nVidia GPU will never see a part with all the SPs enabled. He did it with G80(and of course the 8800GTX/Ultra had all the SPs enabled), he did it with G92(and of course the 9800GTX/8800GTS512 had all the SPs enabled), and he did it with G200(and of course the GTX280 had all the SPs enabled). He has once again done it with Fermi, and we will have to see what comes in the future, but I'm not going to be surprised if we see a GTX490(or whatever) with all 512 SPs enabled.

But really this thread isn't about Charlie, so lets try to keep discussion about him out of it and focus on the topic.

+1 on the bold bit, newtekie1. I'm sure you're right on your Charlie observations too, so let's talk about over PMs (might blab ya later, going out now. )

I think a 512 shader Fermi will come out, but it will be on a smaller process for sure, or it will be too far out of spec for heat and power - the 480 version is bad enough. It seems to me that the Fermi architecture is ahead of current process technology, which is why the GTX 480 has so many compromises as a complete product. Some people just think the Fermi architecture is crap, but I don't think it is.

 

[the54thvoid]

Surely if you're going to talk about Charlie's articles (whether for or against) people should be putting up links to said articles. It's not helpful to say if he was right or worng about something and not quote him through a link. If we want the truth, post links - i thought that was the TPU way?

Anyhow, a 512 core Fermi? - not on this process- wait for the next shrink but like folk have alluded to already, it may be a revised design by then. It's taken 6 months of fiddling to produce a barely thermally acceptable card. Sure you can water cool it but then it's even more expensive and still draws a heap of power.

I think NV will be heads down on making something commercially more viable. Fermi isn't a bad card as such but it is woefully innefficient at what it does (power versus fps). Forget 512 cores and move on. They can shrink it, tweak it, spin it and sell it. And ATI will be doing the same with Southern Islands.

 

[qubit]

Surely if you're going to talk about Charlie's articles (whether for or against) people should be putting up links to said articles. It's not helpful to say if he was right or worng about something and not quote him through a link. If we want the truth, post links - i thought that was the TPU way?

If that were the topic of this thread, then yes. But it was just a few comments, that shouldn't really be on here, so it doesn't matter. Feel free to start a Charlie thread if you like and we'll all pile in!

On the Fermi architecture, I think that the chip is ahead of its time and the process tech can't quite keep up with it, hence a product full of compromises, as I've said in an earlier post here. It looks like they've gone all out on DX11 tessellation performance, which has taken a lot of silicon to implement.

 

[Mr McC]

I don't understand. They already have the performance crown for a single gpu card, at least on paper: how many people actually have a Fermi installed in their system? I don't think anybody's complaining about the performance of the 480, or its closest competitor, the 5870, so adding more performance, be it in the form of shaders or otherwise, does little to address most people's concerns and problems with Nvidia's new cards, namely heat, noise, power consumption, availability and price.

 

[newtekie1]

I think a 512 shader Fermi will come out, but it will be on a smaller process for sure, or it will be too far out of spec for heat and power - the 480 version is bad enough. It seems to me that the Fermi architecture is ahead of current process technology, which is why the GTX 480 has so many compromises as a complete product. Some people just think the Fermi architecture is crap, but I don't think it is.

I don't know what is up with the GTX480 power consumption and heat output to be honest. It doesn't make sense that it consumes 75+w more for just one cluster of SPs and a memory controller.

I'd actually like to see a GTX400 part with all 512 SPs, but with the 320-Bit memory bus.

 

[qubit]

I don't know what is up with the GTX480 power consumption and heat output to be honest. It doesn't make sense that it consumes 75+w more for just one cluster of SPs and a memory controller.

I'd actually like to see a GTX400 part with all 512 SPs, but with the 320-Bit memory bus.

Awe shoot, I'm supposed to be going out, but TPU is too compelling!

Anyway, I'd love to see a 512 SP Fermi with a 512-bit bus. I'd love to know what the theoretical performance of such a card would be (power draw and heat notwithstanding). Heck, how fast would it be with a 1024-bit bus!

Now, I really must run...

 

[extrasalty]

I think soon for nVidia means when they switch to the next process node. Too bad ATI will be there first.

 

[Fourstaff]

512 cores will only happen when they finalize the branding scheme of GTX5xx

maybe its going to be GTX485?

 

[newtekie1]

On the Fermi architecture, I think that the chip is ahead of its time and the process tech can't quite keep up with it, hence a product full of compromises, as I've said in an earlier post here. It looks like they've gone all out on DX11 tessellation performance, which has taken a lot of silicon to implement.

I think a big part of it isn't just the DX11 tessellation implementation, but also the GPGPU push, and a lot of things that were included in the die to implement that. The biggest thing was probably the L2 cache. Which in most of the drawings of the core doesn't look to take up that much space as seen here:

You see that and you don't really get the impression that the L2 is that big of a deal. It doesn't look like it takes up all that much space. But when you look at the real die shot:

You really get an idea of how much space the L2 in the middle really takes up on the die. L2, even on CPUs, is massive. It usually takes up more room on the die than all the processing parts.

I think soon for nVidia means when they switch to the next process node. Too bad ATI will be there first.

All reports point to ATi's next generation stick with 40nm. Globalfoundries' won't be read with 28nm until at least 2011, and probably not until at least the second half before we start to see parts based on 28nm, which means ATi's next generation slated for late this year is still going to be 40nm.

 

[the54thvoid]

Someone earlier said 'they have the single GPU crown' (GTX 480). But now the OC'd ATI production models are coming out, that crown is slipping too. In fact, as far as this review goes http://www.hardwareheaven.com/reviews.php?reviewid=965, they've already lost it...

and if you're thinking it's a ATI biased review, the actual GTX 480 review was very favourable and did give it 9/10 overall (with 10/10 for design and performance)