GF100 evaluation thread

I remember one b3d member ebaying ex review cards...

 

Who? That's pretty much a faux pas, really.

 

Sorry cant remember It was a while ago, I dont think the seller started the thread (it was about old/rare cards iirc) he just sort of joined in halfway through with ive got a few old ex-review cards i'll stick em on ebay

pretty sure at least one was an engineering sample Ive tried searching on "engineering" but had no luck


**edit have an idea who it was, but could be seriously wrong

 

No. It might make sense if nVidia was more like Intel. But the GPU market is far too fluid for this strategy to work out.

 

That's at least sorta kinda reasonable, but it rests on the assumption that nVidia would risk making a large order before being sure of the yields.

 

2-7% difference isn't something to sneeze about. If the difference between the two chips is in the order of a few % then it does make a significant difference especially when considering the HD 5970 vs say 2* GTX 470 and it becomes even more important in the next generation when the cards are expected to be even faster. I would suggest that most P55 boards do not have the extra NF200 chip installed, am I right?

 

That's what the hot lots were for - taking a mitigated risk so as to get further along in the manufacturing process. If they didn't you'd be seeing the Fermi launch a couple of months from now.

Nvidia did take the risk, and only just made it pay off - why do you think there are no 512 parts? That proves that Nvidia had to compromise and didn't get the chip they wanted. They took the risk, and no 512 parts (along with having to use high voltages with the resultant heat/noise) is what they got. If they had waited for the perfect part, not taken any risks, we'd be seeing a faster, cooler, quieter part with 512 stream processors - but another couple of quarters from now.

470/480 is a compromised part, but it doesn't matter as long as Nvidia can find people to buy them as they are.

 

Letting ATI have the only dx 11 chips out for over 7 months might be one reason why nvidia would take a loss on fermi products. It has money in the bank and would rather bleed some of that off then let at have the defacto hardware standard for a generation.

 

Err yes. How else are they supposed to bench it against future cards and try out new driver improvements?

 

Well I believe Charlie had something in the article along the lines of: you'd have to be incredibly arrogant or stupid to do it, oh hi JHH.

 

2-7% is not much at all. You can always clock P55's PCIe bus to 102-107Mhz. It can do that quite easily. I've seen boards that can do 150Mhz on the PCIe.

Even without considering the OC option (despite all the benefits of the P55 based platforms compared to pretty much any other Intel's chipset) a 7% more performance would give like 3 more frames on a 45fps framerate for example, thus reaching 48fps. No biggy!

I don't know if I understood correctly what you wanted to say in your 5970 vs GTX 470 SLI example, but if you meant the 5970 on a PCIe X16 slot (obviously) and the GTX 470 on 8X+8X compared to 16X+16X then yes I see your point. The GTX 470s would lose that little advantage they may had. It is still small though and especially in the GTX 470 case with all their flaws, personally I wouldn't think it twice. A review site should bench at 16X+16X though as this is the norm.

P55 mobos, even high end ones, do not have a NF200 chip in their default configuration as you correctly state. A user that will start consifering a high end rig, will not stop at the extra 50 euros he has to spend though.

 

The only benches I've seen are from anandtech and they are way biased. Most of the games they used are games known to be optimized for Nvidia. Also, the OpenCL benchmarks they used were coded for Nvidia cards so.... even then, it's not impressive.

 

From what I can work out, the Fermi cards are vaguely competitive with ATI cards at the price they currently are in terms of performance. If they are currently selling at a loss to NVIDIA, then bumping the price up to a level where they make a profit makes them non-competitive in terms of price/performance, which is essentially saying "ATI can do just as well as us for much cheaper".

The alternative is to hold off on launch until yields improve and their costs drop and prolonging the joke of announcements about announcements about announcements in an attempt to save face. Either of these options is an implicit acknowledgement that ATI are managing to beat them in the current market. I imagine that it is worth more to NVIDIA to be able to say "we have a card on the market and are competitive" and to prevent any bad press about falling behind ATI , even if it does cost them profit on the cards.

 

Even if they managed to make a small profit on each GTX480/460, that would probably be insignificant compared to design costs. I suspect the current models are mostly for marketing later derivatives and profit/loss on them is not a real concern. That may change if they can get a large amount of them to the market.

Does anyone know what proportion of previous generation cards sold used the high end die? Or what proportion of Evergreen cards sold have been Cypress?

 

I had not considered this, good point.

It does seem like one good real world scenario especially as the enthusiasts seem to have latched onto the P55 platform as the price/performance king. Unfortunately in addition to bandwidth issues I have to wonder about how it would effect issues like micro-stuttering if the card is being run with less than ideal PCI-E bandwidth. Unfortunately I haven't seen a recent comparison so it would be good to see how the 5970 compares to various SLI/Crossfire setups when this feature is enabled and especially how it might scale in the future as ever faster video cards start to push even the PCI-E 2.0 16x bandwidth.

Theres so much to consider with building a high end rig that you would have to excuse most people for overlooking this kind of issue. It is complicated no matter how you look at it.

 

I think Nvidia expect to make their money in the professional/HPC market, where they can sell almost the same card for $3000-5000. Retail sales of these high-end cards at minimal/no profit, without OEM wins or mainstream derivatives won't get them anywhere. That's why they are looking to the HPC market, though they will have to improve on power and heat if they want to go down that road.

 

Dual purpose

 

Hey thats not trying hard enough. They should overclock it and run Furmark, thats not a realistic grill test!

 

Then they should grill beef and not egg.