NVIDIA GeForce 4XX Series Discussion

[Bo_Fox]

A very interesting read from TechReport:

Inside Fermi's graphics architecture
Insider info and careful speculation take us deep into Nvidia's next GPU


http://techreport.com/articles.x/17815/4

Unless I have misunderstood it badly, they are pretty confident that high-end GeForce Fermi (GTX380) will be at least twice as fast as the GTX285:


IMO they are saying "Look at our HD5870 review, because we have GTX285 SLI results there and that's the kind of performance we expect from Fermi cards". That's what I've been thinking since I saw the white papers, I have no reason to think GTX380 will not be at least twice as fast as GTX285. I had no reason to think HD5870 would not be twice as fast as HD4890 either, but the "failure" of one doesn't imply the failure of the other one, especially when Fermi is a complete redesign and RV870 is the last implementation of an aging architecture that IMHO was never thought to be taken to 1600 SPs*.

*In a "beind the scenes" story about RV770 (I think it was in Anadtech) AMD engineer for RV770 said that RV770 was meant to have 640 SPs, until Ati realised they could put 800 SPs within their die area budget and power envelope, due to the high die area efficiency. Another double up would have put RV870 in 1280 SPs and not 1600. IMO that was the plan when R600 was designed and the reason for RV870 not scaling so well.

Here's a sincere reply this time..

Yep, you're right about the 640 SPs.. but that's not too huge of a difference. 640 to 800 or 1280 to 1600 is 25% more. Perhaps you are right that ATI never planned on doing 1600 shaders with this architecture, but it did an astoundingly amazing job with 800 shaders so ATI wanted to just double everything (including the ROP's).. except for the bandwidth of course.

Ok, back to the topic..

It looks like the Fermi will be 2x as fast as the GTX 285, if not a tiny bit more. I would agree that it's gonna be 2x, but some people think that the Fermi architecture will be mainly optimized for CUDA and not for games. That does not make sense.. perhaps CUDA is being given more emphasis, but Nvidia should in no way ignore us gamers (who make up most of the customers).

GT300's specs looks like it will be:

512 shaders (highly likely)
48 ROP's (likely)
384-bit (likely)
GDDR5 memory (highly likely)

That means the GT300 would have 50% more bandwidth than a GTX 285, which already has more bandwidth than a 5870. It might be even more than 50% if Nvidia uses newer and faster GDDR5 memory.

The Fermi architecture has some other wonderful optimizations... especially for running at 8x AA with little to no performance hit versus 4x AA.. finally!

I'm really excited about this. Hurry up Nvidia!!!

 

[troyrae360]

Its Here

 

[Benetanegia]

Here's a sincere reply this time..

Yep, you're right about the 640 SPs.. but that's not too huge of a difference. 640 to 800 or 1280 to 1600 is 25% more. Perhaps you are right that ATI never planned on doing 1600 shaders with this architecture, but it did an astoundingly amazing job with 800 shaders so ATI wanted to just double everything (including the ROP's).. except for the bandwidth of course.

25% is quite a bit, mind you, and maybe, that's the 25% we need to reach the theoretical preformance, the one we can't explain apart from the 25% derived from drivers+memory bandwidth. 25%+25%= the 50% that the HD5870 lacks.

It looks like the Fermi will be 2x as fast as the GTX 285, if not a tiny bit more. I would agree that it's gonna be 2x, but some people think that the Fermi architecture will be mainly optimized for CUDA and not for games. That does not make sense.. perhaps CUDA is being given more emphasis, but Nvidia should in no way ignore us gamers (who make up most of the customers).

GT300's specs looks like it will be:

512 shaders (highly likely)
48 ROP's (likely)
384-bit (likely)
GDDR5 memory (highly likely)

That means the GT300 would have 50% more bandwidth than a GTX 285, which already has more bandwidth than a 5870. It might be even more than 50% if Nvidia uses newer and faster GDDR5 memory.

The Fermi architecture has some other wonderful optimizations... especially for running at 8x AA with little to no performance hit versus 4x AA.. finally!

I'm really excited about this. Hurry up Nvidia!!!

I agree in that just because they focused more on GPGPU, that doesn't mean it won't be good at graphics. The changes they have made are actually good for graphics too, it's just that they are closer to (or well within) the realm of diminishing returns in the graphics department, but the changes are in no way bad for gaming. Gaming and HPC are very similar in what they need, that's why GPGPU always made sense since the beginning.

People is concerned because they think that the changes will make Fermi graphics cards more expensive and gve nothing in return, but that's because they are not seeing Nvidia as the diversified company that it is right now. Nvidia has made Fermi so that it can be used in all of their products and will try to make profits from all of them. If they have to have lower profits from the consumer GPU market, they will, because they will have good income from the HPC and professional market. When you factor in how much they have saved by going the single-chip-does-it-all route they can afford to sell consumer GPUs at a lower margin. At $3000 every Tesla card, they can very well afford a -$10 margin in 100-200 Geforces and they already have contracts for tens or hundreds of thousands of Teslas (ORLN, Cray...), which means they have already covered the reduction in margins for millions of Geforces.

 

[[I.R.A]_FBi]

anyone rushing out to get a g310?

 

[Binge]

anyone rushing out to get a g310?

Please don't crap this thread like troyrae360.

 

[troyrae360]

Please don't crap this thread like troyrae360.

 

[Binge]

I love you too

 

[troyrae360]

I love you too

Aww thanx

 

[HalfAHertz]

...Nvidia has made Fermi so that it can be used in all of their products and will try to make profits from all of them. If they have to have lower profits from the consumer GPU market, they will, because they will have good income from the HPC and professional market. When you factor in how much they have saved by going the single-chip-does-it-all route they can afford to sell consumer GPUs at a lower margin. At $3000 every Tesla card, they can very well afford a -$10 margin in 100-200 Geforces and they already have contracts for tens or hundreds of thousands of Teslas (ORLN, Cray...), which means they have already covered the reduction in margins for millions of Geforces.

Agreed. You have to bear in mind that they need to exceed in just one of the fields to popularise their solution in all of them. If you have the highest performing gaming card word of mouth will help you sell your product in other circles as well...

 

[[I.R.A]_FBi]

Please don't crap this thread like troyrae360.

Im actually being well behaved

 

[qubit]

G310 - Charlie D tells it like it is, again

NVIDIA IS RENAMING parts once again. Now the 200 series is the 300 series, just like we told you months ago. Nvidia has no new parts and no chance of coming up with any soon, so all it can do is unethically try to snow consumers.

If you recall, the roadmaps we showed you in August had the G210 magically becoming the G310 with the addition of DDR3. If you look at this HP page (originally found at Beyond 3D), you can see that there is a G310, but it doesn't appear to have DDR3 yet. This card has literally nothing new. It is a complete joke, and an underperforming joke at that. I wonder if there is a problem with the memory controller, or just HP being cheap?

Yup, more joke parts from nvidia.

Enjoy the rest of Charlie telling it like it is, here.

Pull yer finger out nvidia!

 

[qubit]

FERMI delayed until 2011??!!

According to Chinese newspaper Digitimes, NVIDIA's next-gen GPUs won't ship until the first quarter of fiscal 2011.

Shit, I really hope this Hexus article isn't true.

They could go bust, let alone be uncompetitive.

 

[Benetanegia]

According to Chinese newspaper Digitimes, NVIDIA's next-gen GPUs won't ship until the first quarter of fiscal 2011.

Shit, I really hope this Hexus article isn't true.

They could go bust, let alone be uncompetitive.

Read...

According to Chinese newspaper Digitimes, NVIDIA's next-gen GPUs won't ship until the first quarter of fiscal 2011. That gives us a window of anytime between January 26th and April 26th, 2010.

 

[[I.R.A]_FBi]

still ... LATE

 

[Benetanegia]

still ... LATE

Not later than what we knew 2 weeks ago and not later than other new architectures in history, i.e R600. This kind of delays are in no way rare or exceptional and there's not too much to make out of it.

 

[mR Yellow]

Well, it's a fail for nVidia. 5-6 months without any major competition for ATi is a failure.

 

[Benetanegia]

Well, it's a fail for nVidia. 5-6 months without any major competition for ATi is a failure.

4 months if it is released in january or first half of february, but if it is a failure for Nvidia is still to be seen. No matter how much you sell in the first quarter of new tech, it matters much more how much you sell in the next 3 quarters and Nvidia is aiming at 3 different markets so it will probably pay off, and with excedents.

If you add the fact that AMD is not selling a lot either, because of the shortage and that the shortge will last some more, Nvidia isn't facing a big fail.

 

[wolf]

All pretty standard, its going to happen time and time again, R600 launched and flopped, it couldn't hold a candle to an 8800GTX let alone ultra, then they spent another cycle making them cheaper, and adding some features, again while Nvidia's cards reigned supreme.

Just cos ATI has a bit of an advantage this time round, people are calling MAJOR FAIL on NVIDIA, well I've got some news, they're still selling butloads of cards guys.

"late" is also relative, sure you could say its later than ATI with DX11 parts, but it looks like they don't even give a rats, they have their sights firmly set on Larrabee, where they are right on schedule.

 

[[I.R.A]_FBi]

All pretty standard, its going to happen time and time again, R600 launched and flopped, it couldn't hold a candle to an 8800GTX let alone ultra, then they spent another cycle making them cheaper, and adding some features, again while Nvidia's cards reigned supreme.

Just cos ATI has a bit of an advantage this time round, people are calling MAJOR FAIL on NVIDIA, well I've got some news, they're still selling butloads of cards guys.

"late" is also relative, sure you could say its later than ATI with DX11 parts, but it looks like they don't even give a rats, they have their sights firmly set on Larrabee, where they are right on schedule.

I think they give a rats, why else the hurry to push g310?

 

[wolf]

G310 hardly competes with a 58xx or 59xx card, they're just spamming the lower market at the moment, something AMD is very good at with CPU's too.

It's a game that all the big names are playing very well at the moment, whether they are 'winning' or 'losing', and just because ATI is ahead in the upper high end gfx segment doesn't mean a complete win for them either, especially with their own 'fails' like upping prices since release and lack of 40nm chips to go around. you may well find a few nvidia chips all over the place that are better bang for buck.

I hear of people bidding upwards of 800USD for a 5970 (yes I know thats not retail), sure it has DX11 and eyefinity, but that is not worth price for performance over a $500 GTX295. and the fact is GTX295's are around to buy, as are many other GT200 and G92 chips still. Cypress is friggen' hard to get your hands on.

 

[r9]

I remember the jump NV did with 8800 over previous generation x19xx and 78xx. Ati needed three generations of GPU to match them 29XX -> 38XX ->48XX(I still wonder how they managed to jump from 320 to 800 shaders. And whole that time NV were sleeping.
I`m not surprised from the current situation. NV were late to DX 10.1 late to 40nm late to DDR5 and the next gen DX 11 is just that 40nm GPU paired with DDR5 mem and updated instructions from DX10 where if you have DX10.1 it is much easy job to update to DX11.
ATI are having problems with 40nm process and they have whole lot smaller core NV job is going to be nightmare.

 

[Benetanegia]

ATI are having problems with 40nm process and they have whole lot smaller core NV job is going to be nightmare.

Ati had problems with RV740 too, a much smaller chip. The roblem is at TSMC not with the chips and under that circunstances the 33% bigger core of Fermi doesn't matter too much.

 

[phanbuey]

Ati had problems with RV740 too, a much smaller chip. The roblem is at TSMC not with the chips and under that circunstances the 33% bigger core of Fermi doesn't matter too much.

Well yeah it does... if its 33% bigger means that there is much more room for error during the mfg process.

Everyone will have yield problems with 40nm, but the smaller cores will still have more yield.

 

[Benetanegia]

Well yeah it does... if its 33% bigger means that there is much more room for error during the mfg process.

Everyone will have yield problems with 40nm, but the smaller cores will still have more yield.

Agreed, but when yields are below 40% overall it doesn't matter too much if you have say a 35% or a 25%. Difference is just a 10%. Remember you pay for the entire waffer, so in each case you lose either 65% or 75% and it doesn't really matter as much as when one is having 85% yields and the other one has 60% although in both cases the relation is the same: one has 40% more working chips. That's how I see it.

 

[Binge]

I think they give a rats, why else the hurry to push g310?

Really? You think they're pushing the GF 310?

I'll gladly wait for the GT series of 300 cards instead of being an idiot.