PS3 RSX=NVxx?

[Unreal]

Well, how does the graphics chip RSX of PS3 relates to nvidia? It is nvidia designed and given another name by Sony or is it designed and manufactured by Sony? In the first case that nvidia is behind the chip what is its nearest, in terms of perfomance, PC-GPU? Something in the G7x family(e.g Geforce 7600)?

 

[Rolf N]

RSX has 24 pixel pipes and a 128 bit memory bus which puts it somewhere in between the Geforce 7600 and Geforce 7800 architctures. Due to the narrow local memory bus it falls far short of the 7800's bandwidth of course, but likewise calling it a Geforce 7600 (which has only 12 pixel pipes instead of 24) is harshly downplaying its capabilities.

I think Pete said somewhere the individual pipes were 7800ish instead of 7900ish (less flexible/powerful ALU per pipe; MADD+MUL instead of two full MADDs per clock, or something fairly close). That means NVIDIA's closest codename match is [strike]G70 aka [/strike]NV47.

Not sure about the maker of the chip right now, but I believe it's an IP deal where Sony holds the recipe and orders the chips at an external fab (TSMC?) themselves. Someone correct this if wrong please.

edit: got it.

I believe Sony or Nvidia announced RSX would be running at 550MHz with 700MHz DDR RAM, but there's been some recent talk that it's been downgraded to 500MHz with 650MHz DDR.

A lot of people have mentioned a lot of desktop video cards to compare RSX to, but they're all wrong. RSX, assuming its original core and memory clock speeds and origins ("NV47-based"), is most like a 7800GTX-512, but with somewhat less than half the bandwidth. Of course, that's still not precise, since it doesn't account for the fact that RSX has

1) some extra texture cache,
2) half the ROPs, and
3) (probably most importantly) a tight connection to Cell and Cell's memory in a closed environment.

The fact that RSX was touted as being NV47-based may mean each pixel "pipeline" is just MADD + MUL, rather than the MADD + MADD of the G70 series (7600, 7900, etc.)

So ... G70 and NV47 are not the same thing?

 

[Unreal]

Ok so it is a 7800 GPU with half the ROPs and half the memory width which means (in my opinion) at average its perfomance would probably be between 60 and 70% relative to the perfomance of a 7800GTX.

Sony holds the recipe means that it has bought all the patents from nvidia, or that it is an inhouse design of Sony?

 

[Farid]

Ok so it is a 7800 GPU with half the ROPs and half the memory width which means (in my opinion) at average its perfomance would probably be between 60 and 70% relative to the perfomance of a 7800GTX.

It's more complex than that. RSX is not simply a G70/1, there more to it. It's definitely based on the same architecture though and it also clearly bears a lot of ressemblances with its desktop counterparts, with that said.

Anyway, there are quite a few threads already about RSX in technology forum, try scanning a few of the threads there to learn more about the chip.
We'll try to have a feature, on the site, about RSX in the near future. At least a quick and dirty overview.

 

[Rolf N]

Ok so it is a 7800 GPU with half the ROPs and half the memory width which means (in my opinion) at average its perfomance would probably be between 60 and 70% relative to the perfomance of a 7800GTX.

The thing is it depends on the workload. Focus on bandwidth consumption (blends and AA with just basic shading per pixel) => 7600ish. Focus on internal computational throughput (lots of maths per pixel, many textures per pass) => 7800ish. I don't think there's much value in achieving an even shorter summary than that if you lose so much precision in the process.

Sony holds the recipe means that it has bought all the patents from nvidia, or that it is an inhouse design of Sony?

No, NVIDIA certainly designed the chip, it's their architecture and their IP and Sony pays royalties per chip. However, unlike gfx card makers Sony does not buy the physical chip through NVIDIA, but rather is a direct customer to the fab where they order and pay for the chips themselves, and have the option to renegotiate for process changes, yield improvements, volume rebates etc.

 

[Unreal]

No, NVIDIA certainly designed the chip, it's their architecture and their IP and Sony pays royalties per chip. However, unlike gfx card makers Sony does not buy the physical chip through NVIDIA, but rather is a direct customer to the fab where they order and pay for the chips themselves, and have the option to renegotiate for process changes, yield improvements, volume rebates etc.

This is the point that made me thought that Sony owns the designs since in order to be a direct customer to a chip fab(and order the fab to manufacture a chip in which sony gives its 'RSX' name, Xbox has nvidia's logo printed on the GPU and MainBoard chip, Xbox360 has ATi's Logo on these chips) you ?probably? own the chip designs that you order the fab to produce. It is really strange this situation where nvidia owns the chip design , but sony arranges all the other details such as the ones you mention(plus chip's name) in order to get the chip produced.

 

[almighty]

Ok so it is a 7800 GPU with half the ROPs and half the memory width which means (in my opinion) at average its perfomance would probably be between 60 and 70% relative to the perfomance of a 7800GTX.?

Id say RSX would have a 60-70% performance advantage purely on the basis that RSX will never be CPU bottlenecked and that fact that RSX is in a closed enviroment.

 

[SOcean2forever]

The thing is it depends on the workload. Focus on bandwidth consumption (blends, AA) => 7600ish. Focus on internal computational throughput => 7800ish. I don't think there's much value in achieving the shortest summary possible.

No, NVIDIA certainly designed the chip, it's their architecture and their IP and Sony pays royalties per chip. However, unlike gfx card makers Sony does not buy the physical chip through NVIDIA, but rather is a direct customer to the fab where they order and pay for the chips themselves, and have the option to renegotiate for process changes, yield improvements, volume rebates etc.

Does the fact that RSX has access to both GDDR3 and XDR RAM pools compensate for the bandwidth bottleneck? Obviously the closed environment and CELL are both help with optimization...

 

[swaaye]

....RSX will never be CPU bottlenecked...

I dunno about that.

 

[almighty]

I dunno about that.

Well ok, maybe "never" was abit over the top but if it does get a CPU bottleneck, it wont be very often

 

[nonamer]

RSX has 24 pixel pipes and a 128 bit memory bus which puts it somewhere in between the Geforce 7600 and Geforce 7800 architctures. Due to the narrow local memory bus it falls far short of the 7800's bandwidth of course, but likewise calling it a Geforce 7600 (which has only 12 pixel pipes instead of 24) is harshly downplaying its capabilities.

I think Pete said somewhere the individual pipes were 7800ish instead of 7900ish (less flexible/powerful ALU per pipe; MADD+MUL instead of two full MADDs per clock, or something fairly close). That means NVIDIA's closest codename match is [strike]G70 aka [/strike]NV47.

I'm pretty sure you have the GF6800 and the GF7800/7900 confused. There shouldn't be any difference in ALU configuration between 7800 and 7900.

Not sure about the maker of the chip right now, but I believe it's an IP deal where Sony holds the recipe and orders the chips at an external fab (TSMC?) themselves. Someone correct this if wrong please.

edit: got it.So ... G70 and NV47 are not the same thing?

I'm also pretty sure Sony has their own fabs making these things.

 

[Rangers]

It's a 7900GTX with a 128 bit internal bus and some minor changes.

Dont know why people would want to call it a 7600, it's not. I guess if you think the bus is more important than the rest of the chip, you could, but clearly that's not the case.

 

[jonabbey]

It's a 7900GTX with a 128 bit internal bus and some minor changes.

Dont know why people would want to call it a 7600, it's not. I guess if you think the bus is more important than the rest of the chip, you could, but clearly that's not the case.

What's an internal bus? It has an 128bit interface to the GDDR, it's true, but it has something similar in bandwidth to the XDR and Cell through a separate FlexIO port, giving RSX aggregate bandwidth similar to a G70 with a 256bit interface, as I understand it.

 

[pjbliverpool]

Id say RSX would have a 60-70% performance advantage purely on the basis that RSX will never be CPU bottlenecked and that fact that RSX is in a closed enviroment.

Until we see some evidence that Cell is faster than a good x86 core/dual core for the same kind of in game operations an x86 would be expected to perform when running a game then im very unconvinced of its supposed superiority.

No doubt its better for graphics tasks which would normally be handled by the GPU in a PC and its obviously better video/audio compression than an x86 dual core (lthough not sure about a quad core) however in terms of bottlenecking the GPU in the way you would expect to happen in a PC, then I see no reason to expect Cell to be any better.

Besides, don't you think 60-70% is a bit extreme? Do you really think PC games are CPU bottlecked that much? I think you will find that in most "next gen" games, a 7800 would be by far the bigger bottleneck to performance on a fast dual core system than the CPU.

The closed box argument is certaiinly valid but 60-70% more performance than a part which is quite significantly more powerful on paper (comparing to a 7900GTX) is very optimistic for a closed box advantage IMO. At least when basing the comparison on a well coded game.

 

[Rolf N]

I'm pretty sure you have the GF6800 and the GF7800/7900 confused. There shouldn't be any difference in ALU configuration between 7800 and 7900.

Point taken. So the difference between 7800 and 7900 isn't functional but purely in transistor count/clock headroom/power draw?

I'm also pretty sure Sony has their own fabs making these things.

Yep, I even knew that they have fabs
I wasn't sure they were ready to make RSX yet. If they are, good for them I guess.
At least I had the bus width right

 

[inefficient]

At least I had the bus width right

But then wouldn't some argue that RSX has two 128bit buses?

[size=-2]Not me. I am already bored of this discussion[/size]

 

[Pete]

Edit: Lies! I managed to confuse myself, as Dave pointed out. RSX is 90nm and so most likely G71-based, and that's not exclusive from being NV47-based (as Dave said, and if I can manage to remember something correctly, NV47 was apparently renamed G70 back when the 7800 debuted). So, MADs and 90nm all around for RSX. I compared RSX to 7800GTX-512 based on the core clock and pixel shaders, not manufacturing process. And that's assuming RSX is still 550MHz, as was originally projected, and wasn't slightly downgraded to 500MHz. Basically, ignore me. I'll leave my original post in italics as a painfully funny reminder to myself to think before posting.

Well, original Sony (or Nvidia, I forget) PR slides said RSX was NV47-based. NV47, however, is 130nm, and RSX is supposed to be 90nm. It does seem odd that NV would move G70 to 90nm and not use that for RSX, in which case RSX could be both "NV47-based" and basically a reconfigured G71.

Right, NV47 has a MUL + MAD pixel shader, whereas G70 (and so G71) upped that to MAD + MAD. So, G7x is an upgrade, but I don't know how significant of one. I'm sure someone's clocked a 7800 and a 7900 identically and compared synthetic test #s. Alternatively, 7800GTX (256MB) and 7900GT are pretty evenly clocked.

Again, I say 7800 only b/c of that PS3 PR slide. It may well be 7900. Odds are probably good, given that it's 90nm and so's G71, but I don't know for sure. 7800 or 7900, it's a pretty powerful GPU for a closed system.

 

[Dave Baumann]

Pete - G70 = NV47. At least, thats how RivaTuner and even NVIDIA's dev apps reported the Chip codename.

 

[Pete]

Cripes, I have no idea what I was thinking. For some reason I read zeck's post and thought I myself had said NV47 was like NV4x, not G7x.

Between this and the SSAA slip, I think I'm due for a time-out.

 

[Acert93]

Well, original Sony (or Nvidia, I forget) PR slides said RSX was NV47-based. NV47, however, is 130nm, and RSX is supposed to be 90nm. It does seem odd that NV would move G70 to 90nm and not use that for RSX, in which case RSX could be both "NV47-based" and basically a reconfigured G71.

Again, I say 7800 only b/c of that PS3 PR slide. It may well be 7900. Odds are probably good, given that it's 90nm and so's G71, but I don't know for sure. 7800 or 7900, it's a pretty powerful GPU for a closed system.

As for the process issues, I believe the Sony fabs where RSX are to be made are SOI which, at least accourding to some of the more informed posters here, would not be a trivial change.