The official GT300/Fermi Thread

[Helloworld_98]

GDDR6? What's that?

I thought they had already set the mark for GDDR6, but I guess not.

[LordEC911]

I thought they had already set the mark for GDDR6, but I guess not.

Not standardized yet.

[saaya]

I am very sure that MARS did not exceed 300 W power consumption in PCI SIG's tests. I have no clue about their testing methods, and I wouldn't be surprised if they just ran 3DMark06 and took the average consumption. Or then used their very own testing software and testbed.

3DMark03 Nature at 1280x1024, 6xAA, 16xAF.
Highest single reading during the test.

315W

Maximum: Furmark Stability Test at 1280x1024, 0xAA.

453W(!!!)

http://www.techpowerup.com/reviews/A...5_MARS/27.html

so yes, they did exceed 300W, not by much but they did...
and yes, of course its possible to use 2 8 pin connectors and go for more than 300W... and in reality you dont even need 2 8 pin connectors as highend mainboards can usually supply more than 100W through the slot and 8+6 can deliver more than 225W, its not like the psu will shut down if a card pulls more than 12.5A through a 8pin vga connector... yes, you break the spec... but who really cares? some cards are already pulling more than 100W through the pciE slot, and the spec only allows 75W...

nvidia said gt300 power consumption will be comparable to the current gen... that doesnt really mean anything as there isnt much headroom left to go up... as we all know thermals are holding vgas back...

Not going to happen with 10% yields. They didn't even have enough silicon for their PR event...

according to pcgh, nvidia told them that gt300 will go for a re-spin before itll be sold to retail iirc? but doing a respin in around 4 weeks? is that even possible? hot lots again?

[Calmatory]

So the PEAK was 315, average something like 280 W, and this in a single benchmark run. Hence the card falls to the under 300 W in games.

My point is that there won't be a card which just uses more than 300 W constantly in games. Most probably there won't be a retail card exceeding Mars' power consumption. And as PCI-E standards should be backwards and forwards compatible, the 300 W standard limit isn't going anywhere anytime soon.

So, all in all, this means that GF100 X2 card won't be much more power hog than the current X2 cards.

[ajaidev]

So the PEAK was 315, average something like 280 W, and this in a single benchmark run. Hence the card falls to the under 300 W in games.

My point is that there won't be a card which just uses more than 300 W constantly in games. Most probably there won't be a retail card exceeding Mars' power consumption. And as PCI-E standards should be backwards and forwards compatible, the 300 W standard limit isn't going anywhere anytime soon.

So, all in all, this means that GF100 X2 card won't be much more power hog than the current X2 cards.

Nvidia it self is promoting CPU-GPU software that means the GF100 cores will be used in higher percentage than games. Other than that the GF100 seems to have 8 + 6 instead of 6 + 6, if rumors are to be believed and GF100 does eat around 230W spread around 8 + 6 and PCIe slot the X2 would most likely have 8 + 8 instead of 8 + 6 used in GTX 295 "If a shrunk is not used"

[Calmatory]

Nvidia it self is promoting CPU-GPU software that means the GF100 cores will be used in higher percentage than games. Other than that the GF100 seems to have 8 + 6 instead of 6 + 6, if rumors are to be believed and GF100 does eat around 230W spread around 8 + 6 and PCIe slot the X2 would most likely have 8 + 8 instead of 8 + 6 used in GTX 295 "If a shrunk is not used"

Still the power consumption must remain under 300 W in PCI SIG's internal testing, otherwise the card does not meet PCI-Express standards and hence can not be sold as PCI-Express compicant device. Oh and there is only talk about 6+8 pin in the standard, as far as I know.

[LordEC911]

according to pcgh, nvidia told them that gt300 will go for a re-spin before itll be sold to retail iirc? but doing a respin in around 4 weeks? is that even possible? hot lots again?

Well, I don't know much about hotlots other than how risky they are. Respins usually take 6-8weeks to get back from TSMC, which means they will get it back the middle of Nov.

So, best case for Nvidia if this next spin is good to go and they start ramping production, they might get a handful of wafers ready for the end of the year, like I said before, maybe a couple hundred cards.

Nvidia it self is promoting CPU-GPU software that means the GF100 cores will be used in higher percentage than games. Other than that the GF100 seems to have 8 + 6 instead of 6 + 6, if rumors are to be believed and GF100 does eat around 230W spread around 8 + 6 and PCIe slot the X2 would most likely have 8 + 8 instead of 8 + 6 used in GTX 295 "If a shrunk is not used"

Again... you cannot use an 8+8pin on a "single" GPU, you will not pass PCI SIG certification.

[ajaidev]

Using 8 + 8 does not mean they will use 300w+ avg. The GF100 is suppose to have 8 + 6 and not use the combo to the max, having extra power on tap does not mean the card will use it, but in super intense situations extra power will come in handy "As is with MARS"

[RedBull78]

Nvidia Admits Showing Dummy Fermi Card at GTC.

http://www.xbitlabs.com/news/video/d...r_Q4_2009.html

[saaya]

http://www.bit-tech.net/news/hardwar...-wasn-t-real/1

tim says somebody showed him a picture of what is supposed to be a real fermi card, but he couldnt make out anything cause it was a mobile phone pic and basically more wires than pcb... sorry, i dont buy it...

o check out our brand new fermi card! see! we DO have real next gen cards, and they are working fine!
x huh? thats not a real card!
o yes, this IS a real fermi *cough* card!
x no its not! i hear something moving inside when shaking it!
o ohhh i think you misunderstood me. I said its a real fermi prototype card... the retail cards will come later...
x so you DONT have a real fermi card right now...
o of course we do!
x well you called me here to show them to me, so where is it?
o ohhhh thats too bad, fedex JUST picked them up 5mins ago... im so sorry...
x . . .

[Cybercat]

The biggest difference between ATI and NVIDIA, the way I see it at least, is that ATI is basing his cards on "logic" sp's, which means that if you are able to make and awesome driver, you get and awesome performance, but if you don't...well, you stick with your reals sp's (160 in RV770, 5 ways each so theorically this makes 800sp's). On the other hand, NVIDIA is not that reliant on drivers as their arquitecture is more "beasty": no logical sp's means that you can improve a few algorythms but that's all...

So, ATI has a long way in order to make the RV870 as efficient as RV770. The only way we are seeing that poor scaling is because of that, they have to improve how the sp's work (in order to keep all of them feeded) and how to manage the new AF which is awesome but is showing as a HUGE problem (compared to older gpus in which AF has been nearly free).

That could also explain why there isn't more difference between 5870 and 5850: the more SP's you have, the more you have to work on a driver so we could say that 5870 is gonna get a lot better than 5850 will (and 5850 is already and awesome product nontheless). Within 6 months time I expect 5870 to totally kill 4870X2 performance wise in ALL games/benchmarks.

I agree completely. Just as you said, it's probably a lot harder for ATI to keep newer parts fed the wider their SP engine gets. I actually think their utilization goes down significantly as the SPs go up, resulting in poorer efficiency with each subsequent generation. It's why a GTX 295 with 1.8 TFLOPs is able to beat a 5870 with 2.7 TFLOPs. The amount of FLOPs matters even less than it used to, because it's all in how you use them.

A GF100 might have 1.7 TFLOPs when it gets released, which is only about 60% more than a GTX 285 at 1 TFLOP (more if you're rounding, of course). However, 1/3 of the theoretical FLOP rate in G80-architecture-based cards goes almost entirely unused in real-world situations. The FLOP rate on GF100 is what it can achieve in real-world, and on top of that, it's more efficient than G80/GT200's baseline FLOP rate.

AMD definitely has a ways to go with their drivers, as with a VLIW architecture they rely heavily on their compiler. Has anyone done any testing of AF performance specifically?

[Bodkin]

http://www.bit-tech.net/news/hardwar...-wasn-t-real/1

tim says somebody showed him a picture of what is supposed to be a real fermi card, but he couldnt make out anything cause it was a mobile phone pic and basically more wires than pcb... sorry, i dont buy it...

o check out our brand new fermi card! see! we DO have real next gen cards, and they are working fine!
x huh? thats not a real card!
o yes, this IS a real fermi *cough* card!
x no its not! i hear something moving inside when shaking it!
o ohhh i think you misunderstood me. I said its a real fermi prototype card... the retail cards will come later...
x so you DONT have a real fermi card right now...
o of course we do!
x well you called me here to show them to me, so where is it?
o ohhhh thats too bad, fedex JUST picked them up 5mins ago... im so sorry...
x . . .

Bittech is a very safe source, I dough they would risk reputation on little thing like this

[003]

Still the power consumption must remain under 300 W in PCI SIG's internal testing, otherwise the card does not meet PCI-Express standards and hence can not be sold as PCI-Express compicant device. Oh and there is only talk about 6+8 pin in the standard, as far as I know.

Bullcrap the MARS uses over 450W at peak. As mentioned, the PCI-e slot and the cables can supply way more power than they're rated for. If worse comes to worse, they just won't advertise it as PCI-e compliant

I mean it is THEIR product, just because it doesn't pass PCI SIG testing, how can they prevent them from releasing it? Perhaps they will be paid off by nvidia.

I agree completely. Just as you said, it's probably a lot harder to keep newer parts fed the wider their SP engine gets. I actually think the utilization goes down significantly as the SPs go up, resulting in poorer efficiency with each subsequent generation.

That is a much larger problem for ATI than nvidia.

[Cybercat]

That is a much larger problem for ATI than nvidia.

Er, I should have specified I meant ATI, but yes. I'll make an edit.

[Calmatory]

Bullcrap the MARS uses over 450W at peak. As mentioned, the PCI-e slot and the cables can supply way more power than they're rated for. If worse comes to worse, they just won't advertise it as PCI-e compliant

I mean it is THEIR product, just because it doesn't pass PCI SIG testing, how can they prevent them from releasing it? Perhaps they will be paid off by nvidia.

Are you an idiot OR did you beknowingly miss my post which said

Still the power consumption must remain under 300 W in PCI SIG's internal testing

Also the PCI-E standard says:

A single x16 card may now draw up to 300 W of power up from 225 W

Every PCI-Express device MUST BE certified by PCI SIG, otherwise the device can not be sold as PCI-Express device.

Whatever it is, upcoming Nvidia cards WILL NOT GO OVER 300 W on average. Period.

[saaya]

Bittech is a very safe source, I dough they would risk reputation on little thing like this

oh dont get me wrong, i completely trust tim on this!!!
its just that he said himself, he couldnt really tell much from the picture he saw, and it sounds like he wasnt 100% convinced that the picture he saw was the REAL REAL fermi card... otherwise he wouldnt even have mentioned that it was a blurry cell phone pic of lots of wires coming off a pcb... if hed really believe he had seen fermi in that moment, he would have written just that, that nvidia then DID show him a picture of the actual card...

btw, a good friend apparently managed to get a copy of the picture in question and mailed it to me

behold! this little puppy right here... this is fermi!

Whatever it is, upcoming Nvidia cards WILL NOT GO OVER 300 W on average. Period.

i dont think so either... power aside, the only way to cool 300W or even more is with a 3slot+ heatsink or with water... and thats just too expensive...

[zerazax]

Until you actually figure it out.I went old school and drew it out on graph paper.(The guys with CAD can confirm this.)

With ATI Cypress at 334mm^2(18.2mmx18.2mm plus .8mm border for cutting(19mmx 19mm)) I get 164 canidates per wafer (41 per quarter).

I took Nvidia's smallest estimated measurement floating around the web at 467 mm^2 (21.4mmx21.4.mm plus .6mm for border cuttting(22mmx22mm))30 per quarter . I get 120 canidate per wafer .

Now even with 10% defect rate added only to ATI you still get 148 ATI to Nvidia's prefect 120 canidates per wafer.The number of gets smaller over 22mm x 22mm

So ATI's faulty yield is 23% greater than Nvidia's perfect yield.

I have no doubt Fermi is an awesome card.

This may work for Tesla at $2000 a pop vs Firestream, but Geforce380GTX against Radeon 5870 the math for Nvidia's partners isn't so good.

It'll come down to profit margins if Fermi will survives.
Basicilly if you make Telsa cards you will be okay, selling GeForce will squeeze margins too tight.

My question is what does BFG,eVGA,XFX and Zotec think Fermi will do for their profit margins.

Nvidia's partners may have the world most powerful GPGPU lying at their feet, but if the partners can't profit from it,well?.

That was an example of smallest Nvidia prefect yield against ATI imprefect scenario

The real question is if Nvidia's partners will profit from this beast.
I think Fermi to Geforce will hurt somebody's profit margin.

Don't forget that yields decrease exponentially with size too... not linearly

Either way you look at it, the math favors ATI on price, especially if the experience with 40nm isn't going pleasant for Nvidia

[Calmatory]

Don't forget that yields decrease exponentially with size too... not linearly

Source?

Unless it's based on the fact that when the chip size grows, less chips fit on the wafer, AND when the chip size grows, there is higher chance to have a defect on the chip.

So actually that makes some sense. Defect rate grows linearly, while chips per wafer decrease linearly. As these two add up, yields decrease expotentially. No?

[saaya]

Source?

Unless it's based on the fact that when the chip size grows, less chips fit on the wafer, AND when the chip size grows, there is higher chance to have a defect on the chip.

So actually that makes some sense. Defect rate grows linearly, while chips per wafer decrease linearly. As these two add up, yields decrease expotentially. No?

afaik the amount of "holes" in the wafer is almost constant between a wafer with many small and few big chips... if a small chip that gets you 400 chips per wafer has a yield of 90% that means probably around 40+ defects per wafer (even if a chip gets struck it may still work with some redundant logic disabled or as a cut down version etc)

if you go for a bigger chip that only fits around 100 times on the wafer like gt300, having 40 defects per wafer means you will have a yield of 60%+ plus because the bigger the chip the higher the chance two defects happen in the same chip.

rv870 is 333mm^2 and rumored to have started with yields of ~60% it seems.
im just guessing here, but 333mm^2 should mean they can get around 175 chips per wafer, so that means 105 functional chips which means 70+ defects. gt300 should be around 550mm^2 which means around 100 chips per wafer max, and with 70+ defects, 30+ fully functional chips.

another factor is the bigger your chip, the more wafer space is wasted on the edges, but thats not a huge diference...

wafer costs are around 3000-5000 us$, so 30 chips per wafer = 100-166$ per gpu, pure die costs

for rv870 it should be around 100 chips per wafer so 30-50$ per chip cost...

these numbers are just examples, they arent accurate...
but you can see, for a rough 50% transistor increase of gt300 over rv870, the costs more or less tripple...

[To(V)bo Co(V)bo]

At least talking about supporting technologies and so (the new C/C++ compiled code support, for example). Regarding computing power, I don't have it so clear.

The only example that I have been able to find comparing GPGPU performance in equality of conditions (that includes both running the same code, so it has to be a DirectCompute or OpenCL piece), the HD5870 is pulverizing the GTX285.

And the funniest part is that said example, is an NVIDIA demo of DirectCompute (run on both cards in AnandTech).

To be exact, this one:



But yes, I'm of the opinion that NVIDIA is one generation ahead when talking about GPGPUs technologies. We will see in terms of GPGPU performance when we have something else than this little demo to compare. I think it's a too specific program to make any conclusion from this.

here too is a comparison to crunching on guru.

http://www.guru3d.com/article/radeon...review-test/25

[saaya]

has anybody read the gt300 whitepaper?
i just saw that nvidia said enabling ecc caused a performance drop of around 20% as a result of over bandwidth? :o
I hope desktop gpus will have ecc disabled then!

[Helloworld_98]

wafer costs are around 3000-5000 us$, so 30 chips per wafer = 100-166$ per gpu, pure die costs

for rv870 it should be around 100 chips per wafer so 30-50$ per chip cost...

AMD must be making a huge profit on the 5870's then, 5850's too.

[Chumbucket843]

Source?

Unless it's based on the fact that when the chip size grows, less chips fit on the wafer, AND when the chip size grows, there is higher chance to have a defect on the chip.

So actually that makes some sense. Defect rate grows linearly, while chips per wafer decrease linearly. As these two add up, yields decrease expotentially. No?

wouldnt that be f(x)=2x? their is a lot more to it than that obviously. here is what dr who has to say about designing gpu's:

With all due respect, I don't know if you ever tried to work on a puzzle that have more than 1 billion parts ... this is what GPU are ... so, it is always easy from your screen point of view to write those kind of statements, never the less, I am sure you never even completed a Lego with 100 000 parts ;-)

[003]

At least talking about supporting technologies and so (the new C/C++ compiled code support, for example). Regarding computing power, I don't have it so clear.

The only example that I have been able to find comparing GPGPU performance in equality of conditions (that includes both running the same code, so it has to be a DirectCompute or OpenCL piece), the HD5870 is pulverizing the GTX285.

And the funniest part is that said example, is an NVIDIA demo of DirectCompute (run on both cards in AnandTech).

To be exact, this one:



But yes, I'm of the opinion that NVIDIA is one generation ahead when talking about GPGPUs technologies. We will see in terms of GPGPU performance when we have something else than this little demo to compare. I think it's a too specific program to make any conclusion from this.

1. I love how they leave out RV770, the real competition to GT200, and

2. That is single precision. GT300 will decimate RV870 in double precision.

[LordEC911]

AMD must be making a huge profit on the 5870's then, 5850's too.

With a kit cost of ~$70-80 yeah they are.
Compared to Nvidia's G200 w/ a kit cost ~$120-130 and GF100 which should be upwards of $150, I think I guesstimated just shy of $200.

1. I love how they leave out RV770, the real competition to GT200, and

2. That is single precision. GT300 will decimate RV870 in double precision.

As I have said before, DP is totally dependant on clockspeeds, if they hit their clock targets, they should have ~40% advantage over a 5870, if they have G200 like clocks, that can drop to a ~20% advantage.

Then what if Cypress has more units on it for the 5890 plus a higher clockspeed?
That could make it even closer.