More info about RSX from NVIDIA

[Titanio]

Sorry, wrong quote, it was for Alpha_Spartan, but anyway i think that the MEMEXPORT function in Xenos allows to make in the ALUs the phisics operations and send them to the main memory as vertex data as a new input for the vertex shaders. This way phisics can be done in Xenos without CPU aid.

Oh I see, thanks. Of course the biggest difference between that and the CPU handling that kind of stuff is that it'd be eating into your main rendering power. It also might be a little more difficult to map your physics algorithms to a shader vs C or C++ in a CPU (but someone with more programming experience with shaders may wish to disagree..but I think more "general" power than that in a GPU will still be better for stuff beyond graphics). In some instances, depending on your game and your focus, it may well make sense.

 

[Rockster]

LIR, possible but not likely. And much of what you describe would still be possible using render to texture, albeit less efficient.

 

[dukmahsik]

so is this this thing a beast or what? god, i cant wait til we get some real world output from rsx and xenos.

 

[hasanahmad]

so is this this thing a beast or what? god, i cant wait til we get some real world output from rsx and xenos.

The Bandwidth between the RSX(256mb) and the CPU(256mb) is confusing comparing it to Xenos. The RSX connects to its memory, if the cpu needs to use the rsx memory it needs to go through it and if rsx needs to use the cells memory it needs to go through the Cell. In the case of Xbox 360 the memory is shared between the xenos and xenon (512 mb shared) so there is no loss of efficiency and runtime moving between different rsx and cell infact both can run on the memory simultaneously. In the end the 35 gb/s vs 22.4 might be the case that the xbox 360 performs faster

 

[Titanio]

so is this this thing a beast or what? god, i cant wait til we get some real world output from rsx and xenos.

The Bandwidth between the RSX(256mb) and the CPU(256mb) is confusing comparing it to Xenos. The RSX connects to its memory, if the cpu needs to use the rsx memory it needs to go through it and if rsx needs to use the cells memory it needs to go through the Cell. In the case of Xbox 360 the memory is shared between the xenos and xenon (512 mb shared) so there is no loss of efficiency and runtime moving between different rsx and cell infact both can run on the memory simultaneously. In the end the 35 gb/s vs 22.4 might be the case that the xbox 360 performs faster

X360's CPU accesses memory through the GPU, iirc. I'm not sure if it's more directly connected to the memory controller, but they don't have individual busses to memory.

 

[hasanahmad]

so is this this thing a beast or what? god, i cant wait til we get some real world output from rsx and xenos.

The Bandwidth between the RSX(256mb) and the CPU(256mb) is confusing comparing it to Xenos. The RSX connects to its memory, if the cpu needs to use the rsx memory it needs to go through it and if rsx needs to use the cells memory it needs to go through the Cell. In the case of Xbox 360 the memory is shared between the xenos and xenon (512 mb shared) so there is no loss of efficiency and runtime moving between different rsx and cell infact both can run on the memory simultaneously. In the end the 35 gb/s vs 22.4 might be the case that the xbox 360 performs faster

X360's CPU accesses memory through the GPU, iirc. I'm not sure if it's more directly connected to the memory controller, but they don't have individual busses to memory.

what im saying is that its more efficient with xbox 360 using shared memory rather than 2 pieces of hardware with seperate memory and would have to access the other piece of hardware for output compared to xbox 360 where the information is processed while sharing the memory and outputting at a higher rate or more efficient rate

 

[Brimstone]

Wow. Looks like G70 and RSX will be pretty good parts. C1 is beginning to look like doodoo.

The R500 is trading off transistors dedicated to shading processing because of the Smart Memory chip fabbed by NEC. It's a different philosophical design than the G70 (RSX). Of course all of the Xbox 360 is that way though. A custom CPU designed for maximum video game performance and a custome GPU that churns the out high fidelity pixels at 720p resolution and sustains a liquid smooth framerate.

 

[Titanio]

what im saying is that its more efficient with xbox 360 using shared memory rather than 2 pieces of hardware with seperate memory and would have to access the other piece of hardware for output compared to xbox 360 where the information is processed while sharing the memory and outputting at a higher rate or more efficient rate

Doesn't HAVE to access the other piece of hardware for output. If anything is latency sensitive, keep it to the more local pool. If it isn't, it doesn't matter where it goes.

I think there's still the question of how X360's CPU accesses memory. It doesn't have its own bus to memory, so it has to go through something else. It may connect more directly to the memory controller however, but I'm not sure if its access is quite as immediate as the GPU's (?) In PS3, Cell and RSX can both have "immediate" access to half the memory if necessary, or less immediate access to the other half.

 

[hasanahmad]

what im saying is that its more efficient with xbox 360 using shared memory rather than 2 pieces of hardware with seperate memory and would have to access the other piece of hardware for output compared to xbox 360 where the information is processed while sharing the memory and outputting at a higher rate or more efficient rate

Doesn't HAVE to access the other piece of hardware for output. If anything is latency sensitive, keep it to the more local pool. If it isn't, it doesn't matter where it goes.

I think there's still the question of how X360's CPU accesses memory. It doesn't have its own bus to memory, so it has to go through something else. It may connect more directly to the memory controller however, but I'm not sure if its access is quite as immediate as the GPU's (?) In PS3, Cell and RSX can both have "immediate" access to half the memory if necessary, or less immediate access to the other half.

I heard that the CPU in xbox 360 does not have a memory controller, its the Xenos which acts as a memory controller thats why you see samsung saying we are making the "Graphics" memory chips

anyone confirm this

 

[mckmas8808]

A good example of TSAA. Look at the tree branches in the background too. Nice.

 

[ralexand]

A good example of TSAA. Look at the tree branches in the background too. Nice.

Cool, does that new AA come for free?

 

[hasanahmad]

A good example of TSAA. Look at the tree branches in the background too. Nice.

Cool, does that new AA come for free?

No AA comes for free. except ofcourse eDram , thats why ATI is touting it so much

 

[ERP]

A good example of TSAA. Look at the tree branches in the background too. Nice.

Cool, does that new AA come for free?

My guess is that they are just enabling supersampling for transparent draw operations (usuall alpha kill stuff that MSAA doesn't work with). This was entirely possible even on NV2X, although not exposed anywhere in the API.

 

[mckmas8808]

Not sure if this was posted or not but here is the G70 PDF

http://www.3dchips.net/test/hard/grafik/G70/PO_GEF_7800_07.pdf

 

[hasanahmad]

so is this this thing a beast or what? god, i cant wait til we get some real world output from rsx and xenos.

The Bandwidth between the RSX(256mb) and the CPU(256mb) is confusing comparing it to Xenos. The RSX connects to its memory, if the cpu needs to use the rsx memory it needs to go through it and if rsx needs to use the cells memory it needs to go through the Cell. In the case of Xbox 360 the memory is shared between the xenos and xenon (512 mb shared) so there is no loss of efficiency and runtime moving between different rsx and cell infact both can run on the memory simultaneously. In the end the 35 gb/s vs 22.4 might be the case that the xbox 360 performs faster

when i brought that up once, Vince called me an idiot

well u were right because xbox has more efficiency than PS3 in terms of enclosed performance closer to theoritically possible while PS3 has more power with less efficiency, its the same case as PS2 all over again.

 

[darkblu]

:?

the average level of intelligence of the posts in this forum is falling drastically. especially when fboys start casting vocal judgements left and right on things they have not clue about whatsoever. in this regard, apparently a user ban does not suffice. so could some kind mod take a look for any 're-emerging' fboys and ban the appropriate ip/nets, pretty please?

 

[ralexand]

A good example of TSAA. Look at the tree branches in the background too. Nice.

Cool, does that new AA come for free?

My guess is that they are just enabling supersampling for transparent draw operations (usuall alpha kill stuff that MSAA doesn't work with). This was entirely possible even on NV2X, although not exposed anywhere in the API.

So this AA only helps for objects rendered with alpha.

 

[Titanio]

well u were right because xbox has more efficiency than PS3 in terms of enclosed performance closer to theoritically possible while PS3 has more power with less efficiency, its the same case as PS2 all over again.

:?

You're opening a can of worms here, IMO. Dopeyfish was making a comment specifically about mem bandwidth, and now you're making statements about system-wide efficiency..no offense, but it seems like you're trying to escalate the debate a little into slightly irrelevant areas for this topic.

Anyway, we've discussed NUMA vs UMA on many occasions, if you want to come up to speed with what's already been debated and what's not, feel free to search. The debate can't really be progressed until we get more info (for example, there seems to be still even uncertainty about XDR latency, so without even that we can't begin to discuss, for example, RSX<->XDR latency).

So this AA only helps for objects rendered with alpha.

Yes, it should have a nice benefit given that these textures weren't AAed before in such a manner, and it was noticeable. You've "normal" AA for everything else as usual (well, hopefully it's not simply "as usual", I'm guessing there are improvements there too).

 

[ralexand]

Thanks for the info, Titanio.

Do you know if speedtree uses alpha to render parts of their trees or is it all geometry?

 

[Titanio]

Do you know if speedtree uses alpha to render parts of their trees or is it all geometry?

I think they use billboarded textures (for the leaves), but I wouldn't be surprised if there's an option to go with all geometry. Not sure though. There's probably a performance tradeoff with the latter too, not sure how viable it'd be generally.