The GT300/Fermi Thread

[Nedjo]

Mmmm haven't followed this thread for like a few days, any changes? Do we know anything besides the fact that they should be shown at CES Las Vegas?

do we know it'll be shown @ CES?

[prava]

I thought DarthBeavis said something about working on a mod for some stand @ CES containing Fermi? At least that is what I understood, let me check it...

[annihilat0r]

sorry for the new thread about the ExPreview article boiz, had forgot about this thread, glad it got merged. On the other hand;

[Piotrsama]

Any reason they don't publish the SP performance of the new products?

EDIT: Found this:
2Q 2010: C2050 3GB - $2499 - 1040 GFLOPS / 520 GFLOPS DP
3Q 2010: C2070 6GB - $3999 - 1260 GFLOPS / 630 GFLOPS DP

The SP performance isn't that much better than the older generation.

[Manicdan]

Any reason they don't publish the SP performance of the new products?

EDIT: Found this:
2Q 2010: C2050 3GB - $2499 - 1040 GFLOPS / 520 GFLOPS DP
3Q 2010: C2070 6GB - $3999 - 1260 GFLOPS / 630 GFLOPS DP

The SP performance isn't that much better than the older generation.

because 8x DP sounds so nice they didnt want to scare off with 2x SP being listed anywhere near it.

[trinibwoy]

because 8x DP sounds so nice they didnt want to scare off with 2x SP being listed anywhere near it.

Yep, that's the side effect of advertising GT200 as 3 flops per clock. In real workloads Fermi is probably a good deal faster than GT200 even in SP but due to their aggressive marketing of GT200 capabilities Fermi looks like a mediocre upgrade on paper.

[Lightman]

So Q2 for Tesla
One can hope GeForce will be first to market before Tesla ...

I need second 5870, but cheaper than it is now

[JohnZS]

It's Q2 for Quadro as well
It is a lot faster than the FX5800 4GB, but.... Q2 is such a way off!
My only hope is that they are concentrating on flooding the market with consumer cards before they unleash the workstation Quadros and later those odd Tesler coil things...

[Nedjo]

My only hope is that they are concentrating on flooding the market with consumer cards before they unleash the workstation Quadros and later those odd Tesler coil things...

I don't think that we can expect anything close to flood.

Nvidia's plan is to bin O'cing friendly chips for consumer, and those less clockable for Tesla & Quadro lineup.

They need as high as possible clocks for GeForce, but binning takes time, and in low yield situation it's even harder!

That's why we can expect plethora of so called GTX370 ("X360" has much resemblance with XBOX ) cards, and lot less 380's...

[annihilat0r]

I need second 5870, but cheaper than it is now

You need a second 5870? What resolution are you running on, Eyefinity?

[orangekiwii]

I just want more competition. The 5870 isn't a big enough jump over my dying gtx285 and the 5970 doesn't float my boat for the sole reason of OCing throttling. I just want to see something else come out to market...and I'd like it sooner than later as my gtx285 is starting whine and choke even at full fan speed.

[clo007]

So Q2 for Tesla
One can hope GeForce will be first to market before Tesla ...

I need second 5870, but cheaper than it is now

Wishful thinking.

It's Q2 for Quadro as well
It is a lot faster than the FX5800 4GB, but.... Q2 is such a way off!
My only hope is that they are concentrating on flooding the market with consumer cards before they unleash the workstation Quadros and later those odd Tesler coil things...

I still see them releasing the Tesla before the Geforce just because of all the build up of Fermi being the next supercomputer. Nvidia is already in talks with many universities for Fermi based supercomputer nodes. It's a big business just as much as us gamers.

[Nedjo]

Wishful thinking.

I still see them releasing the Tesla before the Geforce just because of all the build up of Fermi being the next supercomputer. Nvidia is already in talks with many universities for Fermi based supercomputer nodes. It's a big business just as much as us gamers.

you have some info to back up what you're writing?

and about NV Super computer business: http://www.semiaccurate.com/2009/12/...supercomputer/

[trinibwoy]

Hmmmm, let me see. Press release from ORNL themselves or unsubstantiated blurb from Charlie. Difficult decision.

[Lightman]

You need a second 5870? What resolution are you running on, Eyefinity?

It's for Crysis

[JohnZS]

Hmmmm, let me see. Press release from ORNL themselves or unsubstantiated blurb from Charlie. Difficult decision.

Everybody knows that Charlie knows his stuff
Tesla and Quadro are coming early Q2 2010, however there is no ETA on consumer parts, my guess is they may come 1 month or 2 BEFORE the Quadro and Tesla parts as the FX4xxx and FX5xxx quadro cards came just after the GT200 consumer cards.

John

[Bodkin]

This is based off just the top of my head, but as telsa is for computing which needs to be perfect it would need perfect dies, therefore a lot of dies wont make scratch and will be binned for consumer, so we should see consumer parts before or at the same time as telsa parts. Does that sound even slightly fesable?

[trinibwoy]

Yes, that's correct. Desktops get the worst chips.

http://www.semiconductor.net/article...fects-full.php

Variation is hurting the company's business, which depends on binning. The normal practice is to bin the best chips to the ultrahigh-performance accounts, devices that hit the mean performance and operating voltage metrics to the notebook market, and slightly underperforming chips to desktops.

[Bodkin]

Also as the desktop chips dont need ECC should the yields on them not be higher?

[highoctane]

I really doubt gt300 is going to be binned for laptop use anytime soon but where there is a will there is a way I guess.

The only thing geforce gt300 will be binned against will be the tesla derivatives which if I remember correctly will come the quarter after desktop release. Considering the volume difference & market size that geforce and tesla units will be distributed one could assume binning should not be too tough even if they picked the best for tesla units.

I also don't see ECC being implemented on the desktop units.

Variation is hurting the company's business, which depends on binning. The normal practice is to bin the best chips to the ultrahigh-performance accounts, devices that hit the mean performance and operating voltage metrics to the notebook market, and slightly underperforming chips to desktops.

[trinibwoy]

Also as the desktop chips dont need ECC should the yields on them not be higher?

Yields would be higher yeah (and not only due to the lack of ECC) but yields aren't as much of an issue for Tesla anyway since volumes there are much, much lower. Even if only 10% of good dies pass Tesla qualification they could probably still meet demand.

[Cybercat]

Yep, that's the side effect of advertising GT200 as 3 flops per clock. In real workloads Fermi is probably a good deal faster than GT200 even in SP but due to their aggressive marketing of GT200 capabilities Fermi looks like a mediocre upgrade on paper.

I suppose that's a lot better than looking great on paper and looking....not-so-great in reality. For all the FX comparisons being thrown around, that's one happy deviation we can be thankful for.

[ajaidev]

I think EDC may be implemented but ECC may not. But EDC/ECC would be quite useful OCing like i have experienced with the 5850.

[Bodkin]

Whats EDC? Is it Error detecting?

[ajaidev]

Yep EDC = error detection code.

With the 5850 i have seen memory slow down after a certain point when the EDC kicks in the best thing is there are almost no artifacts. When i OC the memory too much the benchmark score comes down so i have to lower the speed of the memory and see which delivers the best score but in olden times even when you did this you may have little artifacts and that could be very annoying and sometimes even could cause the system to hang.

With ECC the memory could be OC a bit higher than EDC but not too much since fermi has single bit fix...

because 8x DP sounds so nice they didnt want to scare off with 2x SP being listed anywhere near it.

Fermi does not even have 2x SP its more 1.35x for the top 6gb fermi and 1.25x for the cheaper 3gb fermi.





The 5-20% ECC hit AMD notes is very interesting because if we decrease 18-20% from 768 we end up close to the 630Gflops fig. same goes to the SP score. So can it be that Nvidia's initial presentation of SP/DP scores were without ECC and these are with ECC??

That would off course mean Nvidia is on track with clock speed's and we could expect the 1536 SP/768 DP gflops in the consumer cards??