RSX architecture and cost...

[patsu]

Hypothetically, lets say I'm working on a sports game which has a user controllable instant replay camera The camera often ends up in situations with heavy transparency overdraw (many full screen passes) which we have no control over since its a user controlled camera. Time is frozen in instant replay so I need all detail preserved on all textures (including the transparent textures) because that mode gets the most scrutiny by reviewers and players alike (screenshots will be taken in that mode). We also need replay to progress at 30fps minimum when the user controls time. Do you feel there is enough bandwidth on PS3 to accommodate that kind of scenario?

Is this question applicable to Mike Acton and the PD people too. ? Resistance has forest level (lot's of foliage in Somerset), and players can run anywhere in the map, plus in-game screenshot or spectate. They also declared that they used only RSX for graphics work and left Cell alone (little or no pre-processing ?). Would be interesting to learn more about their technique(s) in the upcoming GDC.

EDIT: nAo is another guy who seem to be able to rethink the approach to fix stuff into RSX. What's missing in all these discussions is the context. The raw hardware numbers are telling/interesting but we are missing the assumptions and the development team constraints to interpret all these discussions accurately. The NDA doesn't freaking help too.

 

[MBDF]

What if the PS3 had 512 MB of XDR and 64 MB of high speed gddr3 or gddr4 simply for the frambuffer... would texturing only from XDR be feasable?

I think it would be a good compromise at first glance because it gives you 64MB more memory, and 512MB minus the OS for textures, which can all reside in one place. Not only that but the 64MB can act as a sort of pseudo EDRAM, plus since it's a small amount of memory which is cheaper you have the ability to go for quality (speed) over quantity (MB's).

 

[babcat]

"We know that Cell BE chip has 10 thermal sensors onboard it alone and a voltage regulator that can clock each and every core dynamically according to temperture, calculation complexity and other organic system demands between 3.2 and 6ghz! We know this, because of thermal imaging IBM has performed and how F@H Work Units vary their use of Cell on the PS3 (also 1st hybrid Cell was clocked to 5.6ghz before the on chip voltage regulator popped)!

Question: Since RSX has been said to be able to more than keep up with this variable FLOPS production of Cell, while having the capability to rasterize 4k (4096×2160)to a screen (capable, but we would need two HDMI 1.3 ports to carry that bandwidth off the FLEXiO to a Sony’s 4k SXRD projector). What the hell is the RSX, if it can do that on both the PS3 and Cell/RSX combo Computer Board??? No way is it simply a 7800 chip!!!

Let’s pull the lid off this bugger now! You have 256mgs EVRAM. What kind of bus is cooking code across an embeded setup under that cap? Why haven’t we seen a cross section or blowup pic of the RSX Silicon? Why does it cost more than both BluRay Player and Cell BE to build? Every single other chip in the world has a pic of it on the web bare naked. Including PS2 chips, Cell, Intel’s, AMD’s, Nvidia’s, ATI’s and even Xenos, so where in the hell is a pic and true specs for RSX? What does it have a mouse running in a cage to run it or what?

If it is based on the Nvidia 7800, what’s been added? Extra cache more processors or is RSX some kind of Cell Style Hybrid? We know Kutaragi wanted the most powerful system on the planet. Did he get it?

We’re all tired of these nasty secrets, if they were going to use a two Cell system to begin with, why did they decide to pay more than that in cost to build a simple, cheaper 7800 core into the RSX (if that’s all it really is, they juked). It doesn’t add up!

The only way it adds up (if they are truly using a 7800 core) is to build it like the Cell. With that 7800 taking the place of PPE as the core used as a controller along with 4 to 8 SPE vector cores to do the heavy lifting. You then have a traditional GPU core to ween the Devs off onto the SPE’s, similar to Cell with PPC core!

Having everything programmable along with a similar voltage regulator (on Cell) solves the “Upgradability” and projected “10yr Life Cycle” Riddles!

Then as the CoD4 Devs have said “Xenos being more powerful than RSX is a myth” is verified!

Do us gamers a favor, ask them. Get tricky if you have to, let’s find out exactly what’s under the hood of the RSX that can push out 4K!!!"

I tremendously doubt the RSX has 4 to 8 vector cores. However, I am very curious about the FULL specifications of the RSX. I do not know why the RSX is still such a big mystery if it's only a 7800.

If anyone here has any information about the RSX they can share it would be appreciated.

For example, when will "RSX Best Practices" become available?

 

[Shifty Geezer]

If anyone here has any information about the RSX they can share it would be appreciated.

It's not a flippin' mystery! There's no secret magic sauce. It's a '7800' in implementation. If you're unwilling to accept that, fine, but stop wasting forum time by regurgitating the same question to which you've received the right answer enough times for you to have got the message by now if you were willing to listen to people.

That quote of deepbrown's is classic conspiracy theorist fantasy. It 'doesn't add up' only according to a bizarre and convoluted sum that the conspirator is using to work out the inner workings. Ordinary reasoning shows us it adds up nicely and fits in with everything developers have been saying.

 

[babcat]

I've been convinced for a while that it's basically a 7800. However, there are differences. For example, the post transform cache is larger and also the texture cache. It would be FASCINATING to learn more about these differences.

 

[archangelmorph]

What if the PS3 had 512 MB of XDR and 64 MB of high speed gddr3 or gddr4 simply for the frambuffer... would texturing only from XDR be feasable?

I think it would be a good compromise at first glance because it gives you 64MB more memory, and 512MB minus the OS for textures, which can all reside in one place. Not only that but the 64MB can act as a sort of pseudo EDRAM, plus since it's a small amount of memory which is cheaper you have the ability to go for quality (speed) over quantity (MB's).

Surely it's beneficial to sit your framebuffer in XDR considering you get much better bandwidth to both the CELL & RSX than if it were put in VRAM?

I was under the impression that CELL read/write speeds from GDDR3 was quite bad & so in order to get around it you would sit your frame buffer in XDR so that you get better performance (for things like CELL-based post-processing for example..)

 

[Betanumerical]

I was under the impression that CELL read/write speeds from GDDR3 was quite bad & so in order to get around it you would sit your frame buffer in XDR so that you get better performance (for things like CELL-based post-processing for example..)

Cells write speed to the GDDR3 is horrible something fierce (6/8mb/s) but the read speed is good, around 22.4gb/s.

I could be wrong and mixing my numbers up though.

 

[-tkf-]

Ok sorry to bring this up again - Saw this odd reply on the Uncharted interview...and wondered what you guys thought of it.

The only interesting part is about the chips without a lid on. I don´t recall us seeing the Xenos torn apart or the XeCPU.

 

[betan]

Cells write speed to the GDDR3 is horrible something fierce (6/8mb/s) but the read speed is good, around 22.4gb/s.

I could be wrong and mixing my numbers up though.

cell loads the textures to VRAM, it wouldn't make sense to have a slow bus in that direction.

 

[nAo]

Cells write speed to the GDDR3 is horrible something fierce (6/8mb/s) but the read speed is good, around 22.4gb/s.

I could be wrong and mixing my numbers up though.

Cell is fast at writing to vram and slow at reading from it.

 

[Arwin]

Cells write speed to the GDDR3 is horrible something fierce (6/8mb/s) but the read speed is good, around 22.4gb/s.

I could be wrong and mixing my numbers up though.

Yes, you've mixed them up. Cell writes fast to GDDR3, RSX writes fast to GDDR3 and XDR. The only slow bit is Cell reading from GDDR3, and this is normal. The whole point of GDDR3 is that it's fast in exclusive mode to the RSX. XDR is the kind of memory that is made (in combination with FlexIO no doubt) to be fast in a 'multi-tasking' environment.

(hopefully now I haven't mixed anything up)

nAo, now that you're here, still curious about your response to joker's question above.

 

[Betanumerical]

Cell is fast at writing to vram and slow at reading from it.

So I had my numbers the wrong way around.

 

[nAo]

So I had my numbers the wrong way around.

even the numbers are wrong

 

[Zeross]

 

[Betanumerical]

even the numbers are wrong

How right, It seems I read the table in the thread (I saw it a while ago) wrong.

 

[StefanS]

that was before the downclock of RSX from 550 Mhz to 500 Mhz. So the actual numbers should be a bit smaller.

 

[nAo]

Because Pixel data is the most bandwidth intensive operation of a GPU (texturing comes second) and virtually any pixel operations are pushed off to the EDRAM for 360. Its bascially the EDRAM that makes it feasible to use GDDR3 as system RAM.

Too bad I don't have three eyebrows otherwise I'd raise all of them
Unless you're talking about old gen games, modern games that render stuff with many texture layers + complex post processing effects won't consume more frame buffer bandwidth than texture bandwidth.
When your average shader length is 50+ clock cycles (quite typical number..) you're lucky if you write a single pixel to the frame buffer every 3 or 4 cycles.
EDRAM is really useful 'only' for fillrate intensive rendering.

 

[MBDF]

Is it at all feasible to store all textures in just XDR and read from there?

 

[Arwin]

Is it at all feasible to store all textures in just XDR and read from there?

From earlier discussion, one of the few custom aspects of the RSX is that there are longer texture buffers to accomodate the longer latency when reading textures from XDR. I wouldn't expect a situation where all textures are stored in XDR typical, but certainly mixed use of textures from GDDR3 and XDR should be fairly trivial (ianad).

 

[Fafalada]

The camera often ends up in situations with heavy transparency overdraw (many full screen passes)

Hypothetically, if they are actual full-screen passes (as opposed to close to screen size arbitrary shaped polygons), they don't exactly have to run off a GPU, and hence don't need to be subject to same bandwith parameters

But anyway, I am guessing, you'd have to quantify the scenario a bit more. What exactly you define as "many fullscreen passes". Back on PS2 that could mean up to 50 per frame