(Simple?) questions about Xbox 360 and PS3 bandwidths

[marconelly!]

I hope this will be a simple one.

Let's ignore for a second the Xenos' internal EDRAM bandwidth (256GB/s) and the bandwidth of communication of Xenos' parent board with the daughter board (32/16GB/s).

I'm interested at console's main memory bandwidths. How much bandwidth is there for Xenos' to read/write to/from 512MB of unified memory, and how much bandwidth XeCPU has for this same task? Can they both access memory at the same time or does each have to wait until the other one is done? Can either of them read and write to this memory at the same time?

Simillar questions for PS3. How much bandwidth Cell has for reads/writes, and can it do reads and writes at the same time? Same for RSX. Also does Cell's high clocked XDRAM has any advantage over GDDR3? Can RSX access it's GDDR3 and XDR at the same time, and does that limit Cell's access to XDR? Does RSX accessing two memory pools at the same time technically doubles it's bandwidth?

Also, does Cell being able to access it's memory pool independently of RSX accessing it's own memory pool, and both having their separate bandwidths, in some way make this setup less or more favorable to Xbox 360's unified memory solution where I assume bandwidths have to be shared between CPU/GPU but all the data is in the same space?

 

[ninelven]

erased

 

[V3]

I'm interested at console's main memory bandwidths. How much bandwidth is there for Xenos' to read/write to/from 512MB of unified memory, and how much bandwidth XeCPU has for this same task? Can they both access memory at the same time or does each have to wait until the other one is done? Can either of them read and write to this memory at the same time?

Xenos has about 22.4 GB/s that is shared with the CPU, Xenos also has 10.8 GB/s bandwidth for CPU to feed it data. And viceversa Xenos can feed CPU data too at 10.8 GB/s

Simillar questions for PS3. How much bandwidth Cell has for reads/writes, and can it do reads and writes at the same time? Same for RSX. Also does Cell's high clocked XDRAM has any advantage over GDDR3? Can RSX access it's GDDR3 and XDR at the same time, and does that limit Cell's access to XDR? Does RSX accessing two memory pools at the same time technically doubles it's bandwidth?

Cell has bandwidth worth of 25 GB/s to XDRAM. The RSX can share this bandwidth. XDRAM doesn't have that much of advantage over GDDR3.

RSX has max bandwidth of 22.4 GB/s + 35 GB/s (20+15GB/s). When RSX access XDRAM, it would effect Cell surely. But Cell has quite abit of local memory and cache that should keep it happy.

Also, does Cell being able to access it's memory pool independently of RSX accessing it's own memory pool, and both having their separate bandwidths, in some way make this setup less or more favorable to Xbox 360's unified memory solution where I assume bandwidths have to be shared between CPU/GPU but all the data is in the same space?

Hmm most RSX bandwidth to its memory going to be use up for frame buffer bandwidth. Unless the RSX has TBDR like efficiency for frame buffer, I don't see any advantage to X360 unified memory solution.

 

[marconelly!]

So RSX basically has quite a bit higher total bandwidth to it's GDDR3 + XDR memory pools, but it also needs more of such bandwidth as it's storing it's frame buffers there, while Xenos sends the frame buffers to it's daughterboard EDRAM through the dedicated 30GB/s connection with it, thus not taxing the main ram with frame buffer operations?

Why then was someone here saying that RSX should have higher fillrate for HDR and/or better FP precision for framerate blending?

Xenos also has 10.8 GB/s bandwidth for CPU to feed it data. And viceversa Xenos can feed CPU data too at 10.8 GB/s

I see. How much of CPU <-> GPU bandwidth is there between Cell and RSX, and what is the main importance of this particulal bandwidth?

Also an unrelated question, is it really true that RSX can execute literally 2x more shader ops per sec compared to Xenos (100Gops/s vs. 48Gops/s) and render 2x more polygons (1Tp/sec vs. 500Mp/s). Are these numbers actually true or just some fuzzy math, and if they are true, where does such big advantage come from?

And another completely unrelated question - will either of these GPUs be able to efficiently perform displacement mapping? I have not see that feature mentioned anywhere yet when describing either of them.

 

[Megadrive1988]

I never saw any specs on PS3 saying it could push 1 billion polygons a second compared to the 500 million that Xbox 360 does.

although I wouldn't doubt PS3 could push 1Bpps with minimal lighting, shaders, effects, etc.

 

[V3]

So RSX basically has quite a bit higher total bandwidth to it's GDDR3 + XDR memory pools, but it also needs more of such bandwidth as it's storing it's frame buffers there, while Xenos sends the frame buffers to it's daughterboard EDRAM through the dedicated 30GB/s connection with it, thus not taxing the main ram with frame buffer operations?

yep.

Why then was someone here saying that RSX should have higher fillrate for HDR and/or better FP precision for framerate blending?

Several things RSX is clocked higher. It is assumed to be similar to the PC part, and was design for HDR in mind. Compare to say Xenos with its 4AA as one of its high priority.

I see. How much of CPU <-> GPU bandwidth is there between Cell and RSX, and what is the main importance of this particulal bandwidth?

35 GB/s which compose of 20 GB/s for Cell -> RSX and 15 GB/s for RSX -> Cell.

Also an unrelated question, is it really true that RSX can execute literally 2x more shader ops per sec compared to Xenos (100Gops/s vs. 48Gops/s) and render 2x more polygons (1Tp/sec vs. 500Mp/s). Are these numbers actually true or just some fuzzy math, and if they are true, where does such big advantage come from?

Shader ops doesn't mean much. Its probably true, but it doesn't mean much. As for polygon, RSX is only 550 MHz, triangle setup is generally tied up to clock speed. So unless they beefup that part in RSX, I doubt that claim is true.

And another completely unrelated question - will either of these GPUs be able to efficiently perform displacement mapping? I have not see that feature mentioned anywhere yet when describing either of them.

I am sure they can. Not sure about efficiency though.

 

[marconelly!]

I never saw any specs on PS3 saying it could push 1 billion polygons a second compared to the 500 million that Xbox 360 does.

although I wouldn't doubt PS3 could push 1Bpps with minimal lighting, shaders, effects, etc.

OK, if I remember correctly it was someones educated guess here, based on the leaked G70 specs that placed it's performance at 860M/s, and I guess calculated value for RSX was based on it's faster clock and other factors.


*edit* the actual leaked number for G70 is 860M, not 700M

 

[Xmas]

OK, if I remember correctly it was someones educated guess here, based on the leaked G70 specs that placed it's performance at ~700Mp/s, and I guess calculated value for RSX was based on it's faster clock and other factors.

Both numbers are pretty irrelevant, Xenos is only limited by its 1 vertex/clock output, and RSX (with 8? VS running at 550MHz) could probably be able to transform 1.1 GVerts/s, but triangle setup will limit that number. No one will push that many vertices anyway.

 

[marconelly!]

Both numbers are pretty irrelevant, Xenos is only limited by its 1 vertex/clock output, and RSX (with 8? VS running at 550MHz) could probably be able to transform 1.1 GVerts/s, but triangle setup will limit that number. No one will push that many vertices anyway.

I remember though, it's been said that those 500Mp/s on Xenos were actually supposed to be real life performance with non trivial shaders running. Don't know how relevant that comment was, or does the same thing apply to G70 and RSX. How much would triangle setup limit that 860M number on G70 for example?

*edit* the actual leaked number for G70 is 860M, not 700M