http://www.nvidia.com/object/fermi_architecture.html

Nice link with that video explaining the basics! Just be sure to check its price and power supply requirements before you buy a couple of them ;)

MrS
____________
Scanning for our furry friends since Jan 2002

Fermi launch date:
http://www.theinquirer.net/inquirer/news/1593186/fermi-finally-appear

gdf

Luckily we need to wait only 1 month now to know what is true.

gdf

You will probably be able to OC the Fermi cards, just like the previous cards. However, almost any attempt at doing so will get you past the 300 W. So nVidia don't have to do much to allow this ;)

And it's not like your system suddenly fails if you draw more than that. At least in quality hardware there are always some safety margins built in. A nice example of this is ATIs 2-chip 5970: it draws almost exactly 300 W under normal conditions, but when OCed goes up to ~400 W. It works, but the heat and noise is unpleasant to say the least. And the power supply circuitry is quite challanged by such a load and can become a limiting factor (could be helped by a more expensuve board design).

Is that the "edge"? It depends.. I wouldn't like to draw about twice the power of an ATI for a few 10 % more performance. For me it would have to be at least ~80% more performance for 100% more power. Oh, and don't forget that you can OC the ATIs as well ;)

MrS
____________
Scanning for our furry friends since Jan 2002

NVidia has taken huge steps over the last year towards the low and mid range card market for many reasons, none of them being ATI. This should not be seen as a migration from the top end market, but rather a necessary defence and improvement of their low to mid range market. This is where most GPUs are sold, so its the main battlefield, and NVidia have a bigger threat than ATI, in this war.

NVidias mid range cards such as the GT220 and GT240 offer much lower power usage than previous generations, making these cards very attractive to the occasional gamer, the home media centre users (with their HDMI interface), the office environment, and of course to GPUgrid. The GT240 offers up a similar performance to a 8800GS or 9600GT in out and out GPU processing, but incorporates new features and technologies. Yet, where the 8800GS and 9800GT use around 69W when idle and 105W under high usage, the GT240 uses about 10W idle and up to 69W (typically 60 to 65W) when in high use. As these cards do not require any special power connectors, are quiet and not oversized they fit into more systems and are more ergonomically friendly.

WRT crunching GPUGrid, the GT240 will significantly outperform these and many more powerful, and power hungry, CC1.1 cards (such as a GTS 250, 150W TDP) because the GT240 is CC1.2, and is much, much more reliable (G215 core rather than G92)!

NVidia has also entered the extra-PC market, creating chips for things such as TVs and mobiles. However, this is not just an attempt to stave off Intel’s offensive on the low end GPU market. NVidia have found allies in the form of ARM; Intel also started to take the low end CPU fight to ARM, with the release of Atoms and continued development of small CPUs of limited power. However ARM are not an easy target, not by a long way. ARM sell over 1Billion CPU chips every year, for various devices including just about every mobile on the planet. So NVidia have found very solid support from ARM (especially their computer on a chip Cortex ARM9 product, Tegra 2, which could be making its way to a TV, game console, PDA or netbook near you any time soon).

So Intel’s recent move towards low end CPUs and GPUs, and more importantly their development of single chip CPU and GPUs naturally makes them the common enemy of NVidia and ARM. If NVidia failed to respond to this challenge they would have their market place threatened not just by ATI & AMD but by Intel on the main battlefield (where most cards are).


A few things to note about PCIE technologies:
PCIe operates at 2.5GHz
PCIe 2 operates at 5GHz
PCIe 2.1 operates at 5GHz
PCIe 3 will operate at 8GHz, but this has not yet been ratified – rather, delayed until the second quarter of 2010, and perhaps then some.
It is no surprise that it has not been decided upon, Fermi has not been released!

This highlights the secretive nature of GPU manufacturers, too scared to reveal what they are up to even if it hampers their own sales. It lends weight to the argument that the Fermi cards will be an NVidia loss leader; however we are not privy to the internal cost analysis, just the speculation and you could argue that Fermi is a trail blazing technology; a flagship for newly developed technology, which will lend itself to new mid range cards over the next year or so and set a new bar for performance.

It is worth noting that Intel developed PCIe, so they might want to try to scupper the technology to knobble both ATI and NVidia, saying as Intel don’t have a top end GPU; they might want to deny other GPU manufacturers the opportunity to compete on something they cannot control.

It strikes me as likely, that people who buy a Fermi in a month’s time might have to buy a new motherboard sometime later in the year to fully appreciate the card. Today it would have to work on a PCIe2 (or PCIe2.1) based motherboard, as these are the best that there is, but I would not be surprised if better performance will be brought with a future PCIe3 slotted motherboard:
PCIe2 uses an 8b/10b encoding scheme (at the data transmission layer). This results in a 20% overhead. The PCIe3 will use a 128b/130b encoding system, resulting in a <2% overhead, but this is only one area of improvement.

Interesting but not very relevant to our needs.

NVidia seems to be moving away from the high end gaming market anyway.

Semi Accurate reported the following, with some caution.
GTX480: 512 Shaders @600MHz or 1200MHz high clock (hot clock)
GTX470: 448 Shaders @625MHz or 1250MHz high clock (hot clock)
If correct the GTX470 will be about 10% slower than the GTX480, but cost 40% less (going by another speculative report)!

I am not interested in how it compairs to ATI cards playing games, just the cost and that they don’t crash tasks!

Dont like the sound of only 5000 to 8000 cards being made. You could sell them in one country, and our wee country might be down the list a bit ;)

We started another discussion on Fermi in the ATI thread, so I'm going to reply here:

They desperately wanted to become the clear number 1 again, but currently it looks more like "one bird in the hand would have been better than two on the roof".

If a bird in the hand is worth two in the bush, NVidia are releasing two cards, the GTX 480 and GTX 470, costing $679.99 and $479.99, how much is the bush worth?

You just got to love their consistent naming schemes. But there's still hope the actual products will called 20x0, isn't it?

MrS
____________
Scanning for our furry friends since Jan 2002

NVidia stopped producing their 55nm GTX285 chips. Lets hope they simply use existing improvements in technology, tweak the designs a bit, and move it to 40nm, with DDR5. If Asus can stick two GTX285 chips onto the one card, I think everyone would be able to if the chips were 40nm instead of 55nm. The cards would run cooler and clock better. A 20% improvement on two 1081GFlops (MADD+MUL, not Boinc) might not make it the fastest graphics card out there, but it would still be a top end card and more importantly, sellable and therefore profitable.

Fermi is too fat and NVidia made the mistake of thinking they could redesign and rebuild fabrication equipment as if it was a GPU core design. Bad management to let the project continue as it did. Firmi should have been shelved and NVidia should have been trying to make competitive and profitable cards, rather than an implausible monster card.

The concept of expanding the width of a wafer to incorporate 3.2billion transistors is in itself daft. Perhaps for 28nm technology, but not 40nm. It is about scalability. The wider the chip the greater the heat, and the more difficult it is to manufacturer. Hence the manufacturing problems, and low clocks (600-625MHz). Lets face it, one of my GT240s is factory clocked to 600MHz and it is just about a mid-range card. A GTX280 built on 65nm technology clocks in at 602MHz, from June 2008! A GTS250 clocks at 738MHz. If you go forward in one way but backwards in another, there is no overall progress.

NVidia should have went for something less bulky, and scale could have been achieved through the number of GPU chips on the card. You don’t see Intel or AMD trying to make such jumps between two levels of chip fab technology. Multi chips are the way of the future, not fat chips.

Why not 3 or 4 GTX285 chips at 40nm onto one card with DDR5? Its not as if they would have to rebuild the fabrication works.
Firmi wont be anything until it can be made at 28nm.

I read that the fabrication plants equipment is actually different; they had to redesign some of it.
NVidia has produced at least 2 dual GPU cards, so why stop or leave it at that. The CPU manufacturers have demonstrated the way forward by moving to quad, hex and even 8 core CPUs. This is scaling done correctly! Trying to stick everything on the one massive chip is backwards, more so for GPUs.
http://www.pcgameshardware.com/aid,705523/Intel-Core-i7-980X-Extreme-Edition-6-core-Gulftown-priced/News/
http://www.pcgameshardware.com/aid,705032/AMD-12-core-CPUs-already-on-sale/News/
I was not comparing a GTX 285 to a GTS 250. I was highlighting that a card, almost 2 years old, will have the same clock speed as Firmi, to show that they have not made any progress on this front, and when compared to a GTS250 at 702 MHz, you could argue that Firmi has lost ground in this area and this deflates other advantages of the Firmi cards.

multi-chip approach?

ATI:
Gecube X1650 XT Gemini, HD 3850 X2, HD 3870 X2, HD 2850 X2, 4870 X2...
NVidia:
Geforce 6800 GT Dual, Geforce 6600 dual versions, Geforce 7800 GT Dual, Geforce 7900 GX2, Geforce 9800 GX2, GTX295 and Asus’s limited ed. variant with 2 GTX 285 cores... None?!?
Quantum also made a single board Sli 3dfx Voodoo2.

Are you suggesting that no one buy an Nvidia card based on fermi arch until they shrink it to 28 nm?

Nvidia could have at least produced 40n GTX 285 type chips for a dual GPU board, but they chose to concentrate on low to mid range cards for mass production and keep researching Firmi technologies. To some extent this all make sense. There were no issues with the lesser cards, so they wasted not time in pushing them out, especially to OEMs. The GT200 and GT200b cards on the other hand did have manufacturing issues, enough to deter them from expanding the range. It would have been risky to go to 40nm. Who would really be buying a slightly tweaked 40nm version of a GTX 260 (ie a GTX 285 with shaders switched off, due to production issues)?

But perhaps now that they have familiarised themselves with 40nm chip production (GT 210, 220 and 240) more powerful versions will start to appear, if only to fill the future mid range card line-up for NVidia. Lets face it, a lot of their technologies are reproductions of older ones. So the better late than never, if its profitable, approach may be applied.

The last thing gamers and crunchers need is for a big GPU manufacturer to diminish into insignificance. Even to move completely out of one arena would be bad for competition.

Guys.. you need to differentiate between the GF100 chip and the Fermi architecture!

The GF100 chip is clearly:
- too large for 40 nm
- thereby very difficult to manufacture
- thereby eats so much power that its clock speed has to be throttled, making it slower than expected

However, the design itself is a huge improvement over GT200! Besides DX 11 and various smaller or larger tweaks it brings along a huge increase in geometry power, a vastly improved cache system and it doesn't need to waste extra transistors for special double precision hardware at 1/8th the single precision speed. Instead it uses the current execution units in a clever way and thereby achieves 1/2 the single precision performance in DP, even better than ATI at 2/5th. Fermi even supports IEEE standard DP, the enhanced programming models (in C) etc.

All of this is very handy and you don't need 512 shader processors to use it! All they need to do is to use these features in a chip of a sane size. and since they already worked all of this out (since the first batch of GT100 chips returned last autumn the design must have been ready since last summer) it would be downright stupid not to use it in new chips. And to waste engineering ressources to add features to GT200 and shift it to 40 nm (which was the starting point of the Fermi design anyway).

Like ATI they must first have been informed about the problems with TSMCs 40 nm process back in the beginning of 2008. At that time they could have scaled Fermi down a little.. just remove a few shader cluster, set the transistor budget to not much more than 2 billion and get a chip which would actually have been manufacturable. But apparanently they choose to ignore the warning, whereas ATI choose to be careful.

But now that "the child already fell into the well", as the Germans say, it's mood discussion these issues. What nVidia really has to do is to get a mainstream version of Fermi out of the door.

Oh, and forget about 28 nm: if TSMC has got that much trouble at 40 nm, why would they fare any better at 28, which is a much more demanding target? Sure, they'll try hard not to make the same mistakes again.. but that's far more complicated than rocket science.

MrS
____________
Scanning for our furry friends since Jan 2002

A few more words on why GF100 is too large:

Ever since G80 nVidia designed the comparably simple shader processors to run at high clock speeds. That is a smart move, as it gives you more performance out of the same amount of transistors. The G92 chip is a very good example of this: at the 65 nm process it reached 1.83 GHz at the high end and ~150 W, whereas it could also be used in the 9800GT Green at 1.38 GHz at 75 W. These numbers are acutally from the later 55 nm versions, but these provided just a slight improvement.

This was a good chip, because its power requirements allowed the same silicon to be run as high performance chip and as a energy efficient one (=lower voltages and clocks).

G92 featured "just" 750 million transistors. Then nVidia went from there to 1.4 billion transistors for the GT200 chip at practically the same process node. The chip became so big and power hungry (*) that its clock speeds had to be scaled back considerably. nVidia no longer had the option to build a high performance version. The initial GTX260 with 192 cores cost them at least twice as much as a G92 (double the amount of transistors, actual cost is higher as yield drops with area squared), yet provided just the same single precision performance and about similar performance in games. 128 * 1.83 = 192 * 1.25, simple as that. And even worse: the new chip was not even consuming less power under load, as would have been expected for a "low clock & voltage" efficient version. That's due to the leakage from twice the amount of transistors.

(*) Cooling a GPU with stock coolers becomes difficult (=loud and expensive) at about 150 W and definitely painful at 200 W.

At 55 nm the GT200 eventually reached up to 1.5 GHz, but that's still considerably lower than the previous generation - due to the chip being very big and running into the power limit. ATI on the other hand could drive their 1 billion transistor chip quite hard (1.3 V compared to 1.1 V for nVidia), extract high clock speeds and thereby more performance from cheaper silicon. They loose power efficiency this way, but proved to be more competitve at the end.

That's why nVidia stopped making GT200 - because it's too expensive to produce for what you can sell it. Not because they wouldn't be interested in the high end market any more, that's total BS. Actually the opposite is true: GF100 was designed very agressively to banish ATI from the high end market.

From a full process shrink (130-90-65-45-32 nm etc.) you could traditionally expect doulbe the amount of transistors at comparable power consumption. However, recent shrinks fell somewhat short of this mark. With GF100 nVidia took an already power-limited design (GT200 at 1.4 billion transistors) and stretched this rule quite a bit (GF100 at 3.2 bilion transistors). And had to deal with a process which performs much worse than the already-not-applicable-any-more rule. Put it all together and you run into real trouble. I.e. even if you can fix the yield issues your design will continue to be very power limited. That means you can't use the clock speed potential of your design at all, you get less performance and have to sell at a lower price. At the same time you have to deal with the high production cost of a huge chip. That's why I thought they went mad when I first heard about the 3.2 billion transistors back in september..

MrS
____________
Scanning for our furry friends since Jan 2002

The CPU manufacturers have demonstrated the way forward by moving to quad, hex and even 8 core CPUs. This is scaling done correctly! Trying to stick everything on the one massive chip is backwards, more so for GPUs.

NVidias mid range cards such as the GT220 and GT240 offer much lower power usage than previous generations, making these cards very attractive to the occasional gamer, the home media centre users (with their HDMI interface), the office environment, and of course to GPUgrid. The GT240 offers up a similar performance to a 8800GS or 9600GT in out and out GPU processing, but incorporates new features and technologies. Yet, where the 8800GS and 9800GT use around 69W when idle and 105W under high usage, the GT240 uses about 10W idle and up to 69W (typically 60 to 65W) when in high use. As these cards do not require any special power connectors, are quiet and not oversized they fit into more systems and are more ergonomically friendly.

Regarding the Article from SemiAccurate: the basic facts Charlie states are probably true, but I really disagree with his analysis and conclusions.

He states that:
(1) GF100 went from revision A1 to A2 and is now at A3
(2) A-revisions are usually to fix logic errors in the metal interconnect network
(3) GF100 has huge yield issues
(4) and draws too much power, resulting in low clock speeds, resulting in lower performance
(5) nVidia should use a B revision, i.e. changes to the silicon, to combat (4)
(6) since they didn't they used the wrong tool too adress the problem and are therefore stupid

(1-5) are probably true. But he also states (correctly, according to Anand, AMD and nVidia) that there were serious problems with the reliability of the vias. These are the little metal wires used to connect the transistors. These are part of the metal interconnects. Could it be that nVidia used the A-revisions to fix or at least relieve those problems in order to improve yields? Without further knowledge I'd rather give nVidia the benefit of the doubt here and assume they're not total idiots rather than claiming (6).

(7) nvidia designed GF100 to draw more amperes upon voltage increases compared to Cypress

That's just a complicated way of saying: every transistor consumes power. Increasing the average power draw of every transistor by say a factor of 1.2 (e.g. due a voltage increase) takes a 100 W chip to 120 W, whereas a 200 W chip faces a larger increase by 40 W to 240 W.

(8) Fermi groups shader processors into groups of 32
(9) this is bad because you can only disable them in full groups (to improve yields from bad chips)

G80 and G92 used groups of 16 shaders, whereas GT200 went to 24 shader clusters. Fermi just takes this natural evolution to more shading power per texture units and fixed function hardware one step further. Going with a smaller granularity also means more control logic (shared by all shaders of one group), which means you need more transistors, get a larger and more expensive chip, which runs hotter but may extract more performance from the same number of shaders for certain code.
Charlies words are "This level of granularity is bad, and you have to question why that choice was made in light of the known huge die size." I'd say the choice was made because of the known large die size, not despite of it. Besides: arguing this way would paint AMD in an even worse light, as they can "only" disable shaders in groups of 80 in the 5000 series.

(10) nVidia should redesin the silicon layer to reduce the impact of the transistor variances

I'm not aware of any "magic" design tricks one can use to combat this. Except changing the actual process, but this would be TSMCs business. One could obviously use less transistors (at least one of the measures AMD choose), but that's a chip redesign, not a revision. One could also combat transistor gate leakage by using a thicker gate oxide (resulting in lower clock speeds), but that adresses neither the variances nor the other leakage mechanims.
If anyone can tell me what else could be done I'd happily listen :)

(11) The only way to rescue Fermi is to switch to 28 nm

While I agree that GF100 is just too large for TSMCs 40 nm and will always be very power constrained, I wouldn't count on 28 nm as the magic fairy (as I said somewhere above). They'd rather get the high-end to mainstream variant of the Fermi design out of the door as soon as possible. If they're not pushing this option since a couple of months the term "idiots" would be quite an euphemism..

MrS
____________
Scanning for our furry friends since Jan 2002

AMD did not suddenly jump to 2x6 cores. They moved from single to dual (on one die, unlike Intel who preferred to use glue), then moved to quads, hex and now 12 cores. Along the way they moved from 90nm to 40nm, step by step. In fact Intel will be selling a 6core 32nm CPU soon.

Unfortunately, NVidia tried to cram too much onto one core too early.
As you say Mr.S, an increase in Transistor quantity results in energy loss (leakage in the form of heat), and probably rises to the square too. The more transistors you have, the higher the heat loss and the less you can clock the GPU.

There is only 2 ways round this; thinner wafers, or to break up the total number of transistors into cool manufacturable sizes. I know this may not scale well, but at least it scales.

The 65nm G200 cards had issues with heat, and thus did not clock well - the first GTX260 sp192 (65nm) clocked to just 576MHz, poor compaired to some G92 cores. The GTX 285 managed to clock to 648MHz, about 12.5% faster in itself. An excellent result given the 240:80:32 core config compared to that of the GTX 260s, 192:64:28. This demonstrated that with design improvements and dropping from 65 to 55nm excess leakage could be overcome. But to try to move from 1.4Billion transistors to 3Billion, even on a 40nm core was madness; not least because 40nm was untested technology and a huge leap from 55nm.

They, “bit off more than they could chew”!

If they had designed a 2Billion transistor GPU that could be paired with another on one board and also dropped to 40nm they could have made a very fast card, and one we could have been using for some time - it would have removed some manufacturing problems and vastly increased yields at 40nm, making the project financially viable in itself.

I think they could have simply reproduced a GTX 285 core at 40nm that could have been built into a dual GPU card, called it a GTX 298 or something and it would have yielded at least a 30% performance compared to a GTX 295. An opportunity missed.

I still see a quad GPU card as being a possibility, even more so with Quad Sli emerging. If it is not on any blueprint drafts I would be surprised.

AMD did not suddenly jump to 2x6 cores. They moved from single to dual (on one die, unlike Intel who preferred to use glue), then moved to quads, hex and now 12 cores. Along the way they moved from 90nm to 40nm, step by step. In fact Intel will be selling a 6core 32nm CPU soon.

There is only 2 ways round this; thinner wafers, or to break up the total number of transistors into cool manufacturable sizes. I know this may not scale well, but at least it scales.

I think they could have simply reproduced a GTX 285 core at 40nm that could have been built into a dual GPU card, called it a GTX 298 or something and it would have yielded at least a 30% performance compared to a GTX 295. An opportunity missed.

I still see a quad GPU card as being a possibility, even more so with Quad Sli emerging.

Yeah, I was technically inaccurate to say that the 9800 GT uses 69W when idle; I looked it up and that is what I found! It would be more accurate, albeit excessive, to say that the system usage with the GPU installed and sitting idle can increases by up to 69W, with the GPU perhaps using about 23W (the accurate bit); the additional losses coming from the GPUs power dependent components; the PSU and motherboard (the vague and variable bit). Of course if I say this for the 9800GT I would need to take the same approach with the GT240 (6W idle), and mention that in reality idle power consumption also depends on your motherboards ability to turn power off, your PSU, chipset and motherboards general efficiencies. Oh, and the operating systems power settings, but again excessive for a generalisation.

- XbitLabs compared a GT240 to a 9800GT (p4) and then replace the 9800GT with a GTS250 (p6)! Crunching on GPUGrid, and overclocked by between 12 and 15%, my temps, on 3 different types of GT240s stay around 49degrees C. The only exception is one card that sits a bit too close to a chipset heatsink in an overclocked i7 system. How they got 75degrees from native is beyond me. Even my hot (58deg) cards fan is at 34%.

Anyway, I think the GT240 is a good all round midrange card for its many features, and an excellent card for GPUGrid. For the occasional light gamer, that does not use the system as a media centre or for GPUGrid, you are correct, there are other better options, but these cards are less use here, and I don’t think gamers that don’t crunch are too interested in this GPUGrid forum.

I've got a nice suggestion for nVidia to make better use of these chips: produce a GT250. Take the same chip, give the PCB an extra PCIe 6 pin connector and let it draw 80 - 90 W under load. Current clock speeds of these cards already reach 1.7 Ghz easily, so with a slight voltage bump they'd get to ~1.8 GHz at the power levels I mentioned. That would give them 34% more performance, would take care of the performance debate compared to the older generation and would still be more power efficient than the older cards. nVidia would need a slight board modification, a slightly larger cooler and possible some more MHz on the memory. Overall cost might increase by 10€. Sell it for 20€ more and everyone's happy.

nVidia can already disable shader clusters as they whish. They could sell a GF100 with only 32 shaders active, if this made any sense. If I remember correctly the first cards to employ these techniques were ATIs X800 Pro and nVidias 6800GS. On newer cards they may also have more flexibility regarding the ROPS, memory controllers etc., so I think that's quite good already.

What do you mean by "lack of limitations for a GPU card"? I think it does have severe limitations: the maximum length is defined by ATX cases (we're already pushing the limits here), the height is also limited by the regular cases and power draw and cooling are severely limited. Sure, you could just ask people to plug even more 6 and 8 pin PCIe connectors into the cards, but somehow you also have to remove all that heat.

I could see two or four 40nm GPUs on one board,

Could chips be designed to include spaces, like fire-breakers?

@liveonc:

We'd need at least one PCIe 2.0 16x connection routed to the external box. That would mean it would have to be very close to the PC, but that's no show stopper.

What I could imagine: use a common PCB for 4 (or "x" if you prefer) GPU chips, or separate PCBs for each of them (cheaper, more flexible but you'd have to connect them). Make them approximately quadratic and cover each of them with a large heatsink. There'd probably have to be 2 or 3 heatpipes for the GPU core, whereas memory and power cicuitry could be cooled by the heatsink itself. Then put a 120 or 140 mm on top of each of these "GPU blocks". Leave the sides of the device free enough so that the hot air can flow out. You'd end up with a rather flat device which covers some surface, so ideally you'd place it like a tower parallel to your.. ehm, PC-Tower. It could also be snapped upon the front or backside, so you could easily use 2 of these moduls for one PC and both PCIe links would only have to cover ~10 cm to the motherboard I/O panel. For BOINC this would absolutely rock, whereas for games we might need faster and / or wider PCIe. One couild also replace the individual fans with a single really large one.. but then there'd be a large dead spot beneath its axis. And if only selected GPUs are loaded one couldn't speed fans up selectively. BTW: the fans would have easily accessible dust filters.

Oh wait.. did I just write all of this? Let me call my patent attorney first ;)

MrS
____________
Scanning for our furry friends since Jan 2002

@Robert: the problems with the current rack mounted devices are numerous

- they're meant for professionals and thus feature expensive Teslas
- I don't have a 19" rack at home (and am not planning to buy one)
- the air flow is front to back, with the units designed to be stacked
-> the cooling can't be quiet this way
- a separate Xeon to drive them? I want to crunch with GPUs so that I don't have to buy expensive CPUs!

@liveonc: that cable in the Magma-box looks really long, so it seems like PCIe extension is no big deal :)

@Jeremy: you have to feed data to these chips. 256 bit memory interfaces are fine for single chips, where the entire space around them can be used. 386 and 512 Bit quickly get inconvenient and expensive. In order to place two GT200 class chips closer together you'd have to essentially create a 768 Bit memory interface. The longer and the more curved the wires to the memory chips are, the more problematic it becomes to clock them high. If your wire layout (and some other stuff) is not good you can not even clock the memory as high as the chips themselves are specified.
So without any serious redesign you create quite some trouble by placing two of these complex chips closer together - for zero benefit. You could get some benefit, if there was some high bandwidth, low latency communication interface between the chips. Similar to what AMD planned to include in Cypress as Sideport, but had to scrap to keep the chip "small". But that's a serious redesign. And I suppose they'd rather route this communication interface over a longer distance than to cram the memory chips into an even smaller area.
A single cypress pushes the limits of what can be done with 256 Bit GDDR5 quite hard already: HD4890 was not memory bandwidth limited. HD5870 increased raw power by a factor of 2 and memory bandwidth only by a factor of 1.3 and as a result could use even faster memory. Yet they had to be very careful about the bus to the memory chips, its timing and error detection and correction to reach even these speeds.

MrS
____________
Scanning for our furry friends since Jan 2002

Robert,

I've seen no sign of a BOINC version that can make any use of a GPU from Intel yet.

As for Nvidia giving up the chipset market, isn't that most of their product line now? Why would they be willing to risk making such a large change unless the alternative was bankruptcy?

The anticipated supply limitations for Fermi's will mean that the cards will not be available in Europe for some time.
I expect the cards will first be available in the USA, Canada, Japan and a few hotspots in Asia. They may well make their way to Australia before Europe. No doubt they will arrive in the more affluent parts of Europe first too. So Germany, France and the UK may also see them a few weeks before Spain, Portugal and the more Easterly countries, if they even make it that far. As the numbers are expected to be so few, 5K to 8K, very few cards will make it to GPUGrid crunchers in the near future. So if anyone in the USA or Japan gets one early on, please post the specs, performance and observations as it will help other crunchers decide whether to get one or not (should they ever become available in other areas). If they do turn up, and you want one, grab it quickly (as long as it gets a good review)!

Let us all hope they get really bad reviews, or the gamers will buy them up and very few will make it to GPUGrid!

We need 2 high end competitors! Only that way will new technologies advance progressively. At the time being ATI do not have a competitor! So will an average review change anything? A bad review might be just what they need to joult them back on track. Lets face it, Fermi is a loss leader that will depend on a gimic sales pitch that it outperforms ATI is some irrelivant areas of gaming.
The problem is that anyone can compete at the low end, even Intel! But if there are several low end manufacturers eating up the profit margins of the high end GPU manufacturers, they start to look over their shoulder and perhaps panic, rather than looking to the future.

By the way, nice link RobertMiles.
http://www.brightsideofnews.com/news/2010/2/26/asus-to-introduce-dual-fermi-on-computex-2010.aspx
Asus are going to limit their super Republic of Gamers HD5970 Ares dual GPU card to 1000. If they try a similar trick with Fermi's they may find it difficult to get their hands on enough GPUs!

Guys, I don't get it: why are (some of) you talking as if shrinking GF100 to 28 nm was nVidias only option? Just because Charlie said so? Like I said before: IMO all they need is a Fermi design of a sane size, a little over 2 billion transistors and with high speed 256 bit GDDR5.

Take a look at the GT200 on the GTX285: it's one full technology node behind Cypress and features 50% less transistors, so it should be at least 50% slower. How often do you see Cypress with more than a 50% advantage? It happens, but it's not the average. I'd say that's a good starting point for a Fermi-based mainstream chip of about Cypress' size. nVidia said they'd have something mainstream in summer. I'm still confident they won't screw this one up.

BTW: the 4% yield is probably for fully functional chips. And is probably already outdated. Yields change weekly, especially when the process is not yet mature.

MrS
____________
Scanning for our furry friends since Jan 2002

Some time ago I read 9% yield.
Shrink the transistors and the yield should go up by the square. Fermi's low yield is a result of 3 main issues; a change in casting design (larger), a jump to 40nm, and 3Billion transistors.
If they cast smaller and reduced the transistor count, mainstream cards would flow out of the factories.
My arguement was that if they did this, even with GTX 285 chip design (6 months ago), they would have been producing something competitive. Obviously they should not do this now, because they have Fermi technologies, but I agree they have to shrink the transistor count, Now! In 18months or 2years, they may be able to use 28nm, but by then even Fermi will be dated (mind you, it never stopped NVidia from using dated designs in the past).

NVidia design chips, its what they do, and they are good at it.
Unfortunately they overstreached themselves and started to pursue factory mods to manufacture larger numbers of 40nm GPUs. They tried to kill 2 birds with one stone, and missed l0l
NVidia can, and I think will, produce a modified version of Fermi - with a reduced transistor count (it's what they do), but in my opinion they should have done this earlier.
NVidia have already produced low to mid range products. GT210 & Ion (low and very low), GT 220 (low to mid), GT240 (mid range), but then there is a gap before you reach the End of Line GTX 260's (excluding any older GT92 based GPUs that are still being pushed out). By the way I have tried a GTS 240; it uses G92b (like the GTS 250) and basically does not work well on GPUGrid, ditto for several other mid-high end GT92 based cards. I'm not a gamer.

When you see Asus redign PCBs, add a bigger, and in this case better heatsink and fan, it is no wonder they get much better results. So there is more to be had! They have made 2 vastly more powerful GPUs and both create less noise! Well done, twice.
Tackle Fermi too, please. NVidia need that sort of help.

Of course they're diabling non-functional shaders, just as in all previous high end chips. The GTX470 is rumored to have 448 shaders instead of 512 (2 of 16 blocks disabled).

And as mighty as Intel is, I don't think they've got anything to do with the current Fermi problems. Denying the QPI license yes, but nothing else.

MrS
____________
Scanning for our furry friends since Jan 2002

Thanks! That means:

- 225 - 300 W power consumption, as 2 x 6 pin isn't enough (not very surprising..)
- the early performance numbers are to be taken with a grain of salt, as the drivers are not yet where they should be

MrS
____________
Scanning for our furry friends since Jan 2002

They should be ashamed of stripping this poor little GTX naked and revealing it's private parts to the public!

MrS
____________
Scanning for our furry friends since Jan 2002

Is that 12 X 256MB RAM i see?

Well, that's true.. but didn't anyone consider the feelings of this poor chip? If he's not exhibitionist he'll be seriously ashemed by now and probable ask himself the same question over and over again: "why, oh why did they do it?" Like the Rumpelwichte in Ronja Räubertochter. No wonder he's too distracted to run any world records now!

MrS
____________
Scanning for our furry friends since Jan 2002

I think it is 12 * 128 = 1536 (480) and 10 * 128 = 1280 (470)