Not at the same time. The RSX might, though.The GameMaster said:
Xenos - RSX - What was left out?
- Thread starter blakjedi
- Start date
I thought not. I thought it hadn't the hardware. Only SSAA+HDR, not MSAA+HDRThe GameMaster said:
Though likely, it's still too early to state for a fact that RSX doesn't have one, unless you've got a leaked copy of the official tech-specs of the RSX to hand, which we'd like to see.
Not sure if this is true in conjunction with Cell, unless you're talking RSX vs. Xenos outside of the rest of the system.
Who say's it's not better? It might well be.mckmas8808 said:
Shifty Geezer said:
I'd be interested to see or hear more about potential performance in this area on SPEs too. Displacement mapping is something Kutaragi highlighted as something that could be done very well on Cell.
And yeah, looking at these components out of context is of limited use imo..
I know that, but I was also talking of things such as the "45 degree" tricks on R3xx+ and NV4x+, as well as some of NVIDIA's AF optimizations that tend to be evil when driver-generated, but could be interesting in the hands of a developer. Although TBH, I imagine few devs would even care at all about this (why waste 3 days implementing a perfect solution for this when the GPU manufacturers could do it for you? heh)ERP said:
Uttar
I believe it can be safely assumed that the RSX has a bit more raw power than Xenos, but at the same time is less flexible and thus might not always be as efficient. Both have some features the other doesn't have, but in the end, they're both awesome GPUs. Xenos is just quite a bit more interesting or innovative if you will.mckmas8808 said:
As for real-world performance, it all depends on how well ATI's logic for balancing vertex vs. pixelshaders dynamically works I guess. If its as good as they claim, the performance gap between the two might be much smaller than Sony and Nvidia would like people to believe, maybe even non-existant. We won't really know for a while, it'll probably take 2nd or 3rd generation gaming software before we see anything that really pushes the hardware...
I agree, sounds like a perfect job for FlexIO and one or more SPEs. Any idea just how many polys/sec one SPE might tesselate? Xenos # would suggest it needs 2 cycles to tesselate one poly with its dedicated hardware, how many cycles would a SPE need? Even if it takes 10 or more cycles, one SPE might still match or outclass Xenos' fixed function tesselation unit due to Cell's much higher clockspeed...Fafalada said:
Gollum said:
An important factor to consider is bandwidth. People talk about the SPE's doing all kinds of things, from performing vertex calculations to tesselating poly's as a geometry tesselation unit would.
But passings all that data back and forth will no doubt eat up available bandwidth. There are limits.
So the question is, how much bandwidth? And how well would an SPE really perform at such tasks?
How much vertex data can a SPE realistically be expected to process and send over to the RSX before it becomes unrealistic? Same goes for geometry tesselation.
The best option is ofcourse to do such things in hardware on the GPU rather than to simulate them in software on a distant SPE.
I think the memexport capability of Xenos is incredible... and evens things up a bit with RSX in terms of capabilities.
Maybe I missed the discussion on this but Dave mentioned this (additional) capability of Xenos:
"Additional to the 48 ALU's is specific logic that performs all the pixel shader interpolation calculations which ATI suggests equates to about an extra 33% of pixels shader computational capability."
Has this been accounted for in the computational power of Xenos at all? piecing together the Xenos capability (ATI claims versus MS claims) have been rather difficult to do.
Maybe I missed the discussion on this but Dave mentioned this (additional) capability of Xenos:
"Additional to the 48 ALU's is specific logic that performs all the pixel shader interpolation calculations which ATI suggests equates to about an extra 33% of pixels shader computational capability."
Has this been accounted for in the computational power of Xenos at all? piecing together the Xenos capability (ATI claims versus MS claims) have been rather difficult to do.
There's a direct communication between Cell and RSX of 35 GB/s, which doesn't interfere with the RAM bandwidths, similarly to how Xenos's GPU 'core' has 48 GB/s (32 write, 16 read) with the eDRAM.
In total, the render pipeline for PS3 can use a maximum of 35 GB/s direct communication + 25 GB/s XDR + 22 GB/s DDR = 82 GB/s, though of course there'll be BW gobbled up for program code and other jollies.
SPE's should excel at vertex work. It's been spoken of a lot here how good SPE's are to the point that it was conceived the Cell would do all the vertex work and nVidia's components would just be pixel shaders. The extra programmability of SPE's should make them very versatile for Vertex work and surface synthesis, though of course this takes away from the available power for other things like physics. The presence of tesselation on Xenos releves that pressure from XeCPU, which is where this component metrics end up being meaningless.
In total, the render pipeline for PS3 can use a maximum of 35 GB/s direct communication + 25 GB/s XDR + 22 GB/s DDR = 82 GB/s, though of course there'll be BW gobbled up for program code and other jollies.
SPE's should excel at vertex work. It's been spoken of a lot here how good SPE's are to the point that it was conceived the Cell would do all the vertex work and nVidia's components would just be pixel shaders. The extra programmability of SPE's should make them very versatile for Vertex work and surface synthesis, though of course this takes away from the available power for other things like physics. The presence of tesselation on Xenos releves that pressure from XeCPU, which is where this component metrics end up being meaningless.
I think the MemExport function is over the system bandwidth, is it not? I think it provides CPU<>GPU communication similar to CELL<>RSX (and similar to CPU<>GPU over PCI-E) but doesn't have the bandwidth advantage. I'd say that where MS saved system bandwidth consumption through using eDRAM for backbuffer, Sony saved bandwidth with direct communication. The main difference between approaches is we KNOW backbuffer work is a huge BW demand, by vertex work is an unknown to date. Will that Cell<>RSX BW ever be used to capacity, it will it only ever save a few hundred megabytes a second from system BW?blakjedi said:
I am still HIGHLY sceptical of the SPE's abilities for two reasons.
1. Depending on who you talk to, the SPE's can either be used for everything under the sun, or that they are too specialized and are infact not very useful for a good number of things.
2. People use the SPE's as a catch-all. Don't worry if the RSX can't handle it, the magical SPE's will come to the rescue and can do everything at lighting fast speeds. No concievable limitations or bottlenecks here
There ARE limits, and I find it difficult to put much faith in certian claims when the limits aren't even explored.
How much sense does it REALLY make to have the Cell working on tasks better suited to a GPU?
1. Depending on who you talk to, the SPE's can either be used for everything under the sun, or that they are too specialized and are infact not very useful for a good number of things.
2. People use the SPE's as a catch-all. Don't worry if the RSX can't handle it, the magical SPE's will come to the rescue and can do everything at lighting fast speeds. No concievable limitations or bottlenecks here
There ARE limits, and I find it difficult to put much faith in certian claims when the limits aren't even explored.
How much sense does it REALLY make to have the Cell working on tasks better suited to a GPU?
BenQ said:I am still HIGHLY sceptical of the SPE's abilities for two reasons.
1. Depending on who you talk to, the SPE's can either be used for everything under the sun, or that they are too specialized and are infact not very useful for a good number of things.
2. People use the SPE's as a catch-all. Don't worry if the RSX can't handle it, the magical SPE's will come to the rescue and can do everything at lighting fast speeds. No concievable limitations or bottlenecks here
There ARE limits, and I find it difficult to put much faith in certian claims when the limits aren't even explored.
How much sense does it REALLY make to have the Cell working on tasks better suited to a GPU?
I don't think anyone ever said that there aren't limits?? Cell will help RSX on some things, but there are limits. Kinda natural if you ask me.
london-boy said:BenQ said:I am still HIGHLY sceptical of the SPE's abilities for two reasons.
1. Depending on who you talk to, the SPE's can either be used for everything under the sun, or that they are too specialized and are infact not very useful for a good number of things.
2. People use the SPE's as a catch-all. Don't worry if the RSX can't handle it, the magical SPE's will come to the rescue and can do everything at lighting fast speeds. No concievable limitations or bottlenecks here
There ARE limits, and I find it difficult to put much faith in certian claims when the limits aren't even explored.
How much sense does it REALLY make to have the Cell working on tasks better suited to a GPU?
I don't think anyone ever said that there aren't limits?? Cell will help RSX on some things, but there are limits. Kinda natural if you ask me.
I don't mean to sound grim.
But from what I have read regarding the Cell's usefulness for many things, seems to boil down to PURE conjecture.
I have read speculations that the SPE's will be used for EVERYTHING, such as advanced physics, A.I., AA, processing vertex data, geometry tesselation, Ray tracing ect.
There's just too much hype/expectations based on not enough info. IMO.
I'm very interested to know just what things the SPE's are well equiped to deal with, what things their not, how well they're equiped to deal with them and how complicated ( or not ) it would be to implament such things.
I'd LOVE to read some balanced and informed info about it, but there doesn't seem to be much of that going around. Mostly just people using the SPE's and the Cell in general as a "Catch-all."
BenQ said:
Well it's up to you and all of us to filter the hype from factual info then.
Obviously the SPEs have their uses, and WILL be helping the RSX is some aspects of the graphics, and obviously there will be limits.
Do you think that Sony would have pushed this design so hard if the SPEs were useless pieces of junk? Seeing how cool the VUs were - though a total pain to use properly - i think we can trust Sony that the Cell architecture will have its many interesting uses.
Till we see the games and hear from the devs, this is all we can do really.
Is it really pure conjecture if it's based on experience with more limited variations of similar architecture types?BenQ said:
I mean in the end usefullness and efficiency of unified shaders is based on conjecture too - but many of us actually have reasons why we're excited about their prospect - even if we never seen it in action yet.
Similar threads
- Sticky