3D Capabilities of nForce 4?

[JoshMST]

I think you are getting a bit confused here by Huang's "It cost us 60 million transistors" statement. I think what he basically meant was that they could do a PS 2.0 part with 16 pipes and all of the current tech for about 160 million transistors. Yes, adding SM 3.0 to the mix did increase the amount of transistors per pipeline. However, if you look at the NV3x pipeline, they are also fairly complex, plus they have the extra texturing unit, as well as the other garbage that NV threw in there that didn't need to be. NVIDIA totaly revised their pipeline scheme to be more efficient in math operations (and not as robust in texturing operations). So, mixing and matching those things up led to a decrease in overall transistors per pipeline.

Basically, if NV had just made a PS 2.0 compatible unit that stuck to those specifications, the NV40 would have been a much smaller chip. However, even with SM 3.0 capabilities, the NV4x pipeline is still smaller than the NV3x pipeline (you can do the simple math there cause the NV35 was 125 million transistors and it had 4 pixel units and 3 VS units, and if you brought that architecture to the 16 pipeline design and 6 VS units, that chip would hit 330 million transistors easily).

So, chopping the NV4x down to a 2x1 product for integrated graphics would probably net you around 35 million transistors (lean) up to 45 million transistors (with some bells and whistles). On the 110 nm process, this would be a very affordable chip for the chipset market.

 

[KimB]

I don't get how your coming to that conclusion.
Even if a 2x1 nv4x is 1/8th the pipeline tranistors . There is still the sm3.0 transistors . Which nvidia claimed took 60million.

Which would be divided up among the 16 pipelines of the NV40.

 

[jvd]

I still must disagree . Yes the nv40 is more effcient than the nv3x . But i don't agree that a 2x1 nv4x would be smaller than 2x1 nv3x . The nv3x will indeed be smaller.

I also don't agree with you.

Your talking about a 7.3 decrease in size for removing 14 pipelines ? and vertex shaders (5 vertex shaders i believe correct ? ) Then you go on to tell me that the pipelines in the nv40 are smaller than the nv3x . So how exactly would that work.


Now with the nv3x I'm talking about a 4x decrease . 2 pipelines , 1 tmu and the pool of vertex shaders . Leaving only 1 in there .



but mabye someone here that has more hard figures can post

 

[JoshMST]

Earlier you had mentioned a NV3x part with a 4 x 2 architecture (which is a GF 5200- I guess, since NV is kinda secretive about counting things in there). Still, the FX 5200 as a whole is 45 million transistors, and it had enough taken out of it that it performs pretty poorly as compared to other products that are its size, and NV does not allow it to do FP32 rendering (at least last I heard). Now, a 2 x 1 NV4x derivative would give the standalone FX 5200 a run for its money in most applications (assuming that the NV4x part is running at 300 MHz on an Athlon 64 939 motherboard), plus achieve better PS 2.0 performance using both FP16 and FP32, as well as utilize SM 3.0. Of course being only 2 pipelines, it isn't going to be a barnburner by any stretch of the imagination.

But you would be correct, if NVIDIA decided to chop down the FX 5200 to a 2 x 1 architecture, it would be smaller than a 2 x 1 NV4x.

This of course leads us to another possibility... what if one of the other standalone NV4x chips actually constitutes such a design? Would NVIDIA create such a small NV4x chip for the extreme lowend and mobile market? A 45 million transistor product on the 110 nm process would be incredibly small, and loads of them could be had per wafer. Not only that, but if NVIDIA kept the clock between 300 MHz and 400 MHz, speed bins and yields would probably be pretty good. It certainly would be nice to have something like this totally displace the FX 5200/5500. Or would a 4x2 architecture be more sensible in terms of performance/cost/die size?

 

[KimB]

Well, jvd, just look at the NV43. With 146 million transistors, it doesn't have quite 10% more transistors than the NV35, which has a similar texel rate.

With it only being 10% larger than an NV3x with equivalent texel rate, the NV4x equivalents of the NV34 and NV31 parts may actually be smaller at .11 microns.

 

[Fodder]

Earlier you had mentioned a NV3x part with a 4 x 2 architecture (which is a GF 5200- I guess, since NV is kinda secretive about counting things in there).

Only the 5800 and 59x0 are 4x2.

 

[jvd]

Well, jvd, just look at the NV43. With 146 million transistors, it doesn't have quite 10% more transistors than the NV35, which has a similar texel rate.

With it only being 10% larger than an NV3x with equivalent texel rate, the NV4x equivalents of the NV34 and NV31 parts may actually be smaller at .11 microns.

Right . It has half the vs and half the pipe lines at 146m trans . So thats not even half the amount of tranistors (it would be 110 for half ). Its about 1.5 times less. So if we do that again (for 4 pipe lines ) We have 97.3 million. Once again leaves us 64.8 million transitors . Close to that of what I figured (70 million)

Now Josh points out that the 5200 part is already 45 million transistors it was 4x1 from the beyond3d 3d charts . Not only that but it was on 150nm.

On 110nm this would be a tiny core . It would also most likely offer a higher clock speed and higher fill rate than a 2x1 nv4x .

Now looking at nvida's last igp. It was the nv17 (it did have tnl , have to check friends driver) that was 29m transistors on 150nm . Thats roughly 16 million transistors more than the nv3x .

Now anand is saying ati will have a 9600non pro as thier p4 igp

rv360 was 70million transistors. Though anand doesn't say what they cut out . They do say its a fully intergrated rv360 though. So it lends me to believe its 70 million transitors at 130nm .


Going on that perhaps we will see a nv36 which was 82m at 130nm.
Or a 4x1 nv40 at 97m on 110nm .

I would assume the nv36 would still be much smaller on 110nm and most likely require less power and cooling than a 4x1 nv40 at 97m on 110nm.

I would also assume the yields would be better too.

THey also wouldn't need to design a new aisc or whatever

please correct me if my maths are wrong or if any of my figure are wrong and how to come to the right ones (i'm still learning after all )

 

[Evildeus]

If we we go with the same transistor saving rate, the 2*1 would be around 60-65m transistors. Now if we chop the video part....

 

[KimB]

Right . It has half the vs and half the pipe lines at 146m trans . So thats not even half the amount of tranistors (it would be 110 for half ). Its about 1.5 times less. So if we do that again (for 4 pipe lines ) We have 97.3 million. Once again leaves us 64.8 million transitors . Close to that of what I figured (70 million)

You're forgetting that there's an unknown number of transistors that were added for PCI Express support, which throws off your calculations. Additionally, the NV43 has more than half the VS pipes of the NV40, which again makes those estimations for the number of transistors too large.

Then you have to consider that the low-end NV3x parts had other corners taken in their designs, reducing the number of transistors more. I don't see why this wouldn't happen again.

 

[jvd]

If we we go with the same transistor saving rate, the 2*1 would be around 60-65m transistors. Now if we chop the video part....

what video part ?



Btw 2*1 at 60-6m would still give less performance than the smaller 5200.

You're forgetting that there's an unknown number of transistors that were added for PCI Express support, which throws off your calculations. Additionally, the NV43 has more than half the VS pipes of the NV40, which again makes those estimations for the number of transistors too large.

I thought the nv43 was using the bridge chip ? what would they have on the chip for pci-e ?

Then you have to consider that the low-end NV3x parts had other corners taken in their designs, reducing the number of transistors more. I don't see why this wouldn't happen again.

Well the 5200 is a full feature card. ITs speed is another question. But then again igp is never great performance.

You also have to figure in almost all cases a 2x1 nv43 with 1 vs unit would get creamed by a 9600non pro in ati's igp for the p4s .

I would wager the 5200 would also out perform the 2x1 nv43 with 1 vs unit most of the time too. Esp since the 5200 would be smaller and would most likely reach higher clocks .



If nvidia is able to do it and puts out an acceptable performance igp then its great news. Sm 3.0 at the low end. I just highly doubt they will be able to do it .

I

 

[KimB]

The NV43 is native PCI Express.

While a 4x1 NV4x may have a similar texel rate to the NV34, yes, it'll be larger. But a 2x1 NV4x would almost certainly be smaller.

 

[jvd]

The NV43 is native PCI Express.

While a 4x1 NV4x may have a similar texel rate to the NV34, yes, it'll be larger. But a 2x1 NV4x would almost certainly be smaller.

where did u read that ?

 

[Fodder]

The NV43 is native PCI Express.

where did u read that ?

It's in pretty much every one of the 'previews'. They also mention that it'll use a bridge chip for AGP support. It's in the blurb on the B3D news page if you want to take a look.

Well the 5200 is a full feature card.

From B3D's NV34 preview.

"As we see the 'Intellisample' feature is not listed for GeForce FX 5200 Ultra. Intellisample covers a number of elements, such as Multi-Sampling FSAA and Anisotropic filtering which 5200 does support, however it covers other performance elements, such as colour and Z compression techniques, that have been left out of 5200 in order to keep costs down."

You also have to figure in almost all cases a 2x1 nv43 with 1 vs unit would get creamed by a 9600non pro in ati's igp for the p4s .

Remember what happened last time when ATI said they were putting a 9200 in as their IGP? It turned out to be 2x1 with no VS. I'd like to see a full 9600 onboard, but I think we'll have to what and see.

 

[jvd]

The NV43 is native PCI Express.

where did u read that ?

It's in pretty much every one of the 'previews'. They also mention that it'll use a bridge chip for AGP support.

thanks just gave them a quick read.

Now all we need to know is how many tansistors it takes and how many the video takes .


I find this interesting from toms

NVIDIA's own performance figures for the 6600 GT promise performance on a par with a GeForce 5950 Ultra or Radeon 9800XT, at a lower price

so if the 8x1 6vs unit 6600 compares on par to the last gen parts its replacing i don't see how a 2x1 1vs would compare to a 5200. Esp with the 5200 being much smaller. let alone a nv36 or rv360

 

[Fodder]

so if the 8x1 6vs unit 6600 compares on par to the last gen parts its replacing i don't see how a 2x1 1vs would compare to a 5200. Esp with the 5200 being much smaller. let alone a nv36 or rv360

I don't know about that. If the 8x1 part is on par with the 5950, the 4x1 part should be able to match the 5600/5700, and a 2x1 part might be able to catch the 5200, especially since they'd both be fairly solidly bandwidth limited in IGP form.

 

[jvd]

so if the 8x1 6vs unit 6600 compares on par to the last gen parts its replacing i don't see how a 2x1 1vs would compare to a 5200. Esp with the 5200 being much smaller. let alone a nv36 or rv360

I don't know about that. If the 8x1 part is on par with the 5950, the 4x1 part should be able to match the 5600/5700, and a 2x1 part might be able to catch the 5200, especially since they'd both be fairly solidly bandwidth limited in IGP form.

I guess it would depend.

on dx 9 the 4x1 with 4 vs would most likely match the 5600 mabye 5700. But a 2x1 with only 1 vs i don't think would catch a 5200.

But yes both will become bandwidth limited much quicker i guess.

I dunno i guess we will see . I mean if you think nvidia can put a a nv40 cut down in there and u don't think ati could put in a rv350 in there .... well i dunno where your reasoning is coming in.

Also the fact that at the time nvidia decided to put a geforce 4 mx part in the nforce 2 (basicly a geforce 2 lvl product) and at the time they were putting out the nv30s and they could have put in a geforce 4 based product in but they didn't . Why they would put a nv4x based product now .

And for that matter why ati wouldn't just 2x1 or 4x1 r42x core in the igp.

 

[Fodder]

Because lately ATI seem alot tighter on the purse strings than NVIDIA. Though, I don't think integrated video is important enough an area for JHH to turn on the cashflow and start whipping the engineers.

No one's saying ATI can't do things, just that they might choose not to.

Also the fact that at the time nvidia decided to put a geforce 4 mx part in the nforce 2 (basicly a geforce 2 lvl product) and at the time they were putting out the nv30s

NForce2 was launched in June '02, almost a year before NV3x started arriving in quantity.

 

[jvd]

Because lately ATI seem alot tighter on the purse strings than NVIDIA. Though, I don't think integrated video is important enough an area for JHH to turn on the cashflow and start whipping the engineers.

Also the fact that at the time nvidia decided to put a geforce 4 mx part in the nforce 2 (basicly a geforce 2 lvl product) and at the time they were putting out the nv30s

NForce2 was launched in June '02, almost a year before NV3x started arriving in quantity.

Was it ? I must be thinking of the nforce 2 . Which still makes my point stand . Why did they not upgrade the igp with nforce 2 ?


anyway the only reason i would see them go full out and mabye even loose money for awhie on a nv4x igp would be to get as many sm 3.0 nv4x parts as possible into homes

 

[Fodder]

Was it ? I must be thinking of the nforce 2 . Which still makes my point stand . Why did they not upgrade the igp with nforce 2 ?

Mid 2001: NVIDIA release nForce1 with integrated GF2MX
Mid 2002: NVIDIA release nForce2 with integrated GF4MX
Mid 2003: NVIDIA release low end NV3x parts

 

[Evildeus]

I don't think they are willing to lose money on low end parts. It will be interesting to see if they can deliver a Nv4* part with less than 50m transistors 8)