HD problems in Xbox 360 and PS3 (Zenji Nishikawa article @ Game Watch)
- Thread starter one
- Start date
Mintmaster said:
Sorry I know this is may be off topic but looking at the slide 17 are they only using ~1/12 of the CPU power
(unless the low 5,5 (from ~71 IIRC) gflops number are tied with something else)Also only half of the GPU, would they have room enought to, in a future UE3.x, have a new shadow system that can be good enought to also do 4xAA.
Last edited by a moderator:
It is kind of off-topic, but I read in another thread someone was saying about how dynamic branching and other SM3.0 things will play heavily in Xenos' favour. I am wondering what the justification is for this? Is it simply inference based on the fact that ATI's SM3.0 PC parts are better at it than nVidia's SM3.0 PC parts?
Batch sizes for Xenos are 64 Vertices / Pixels, whereas Pixel shader batch sizes for G71 are 1024 pixels - so the PS branching cost for Xenos is many times smaller; the VS branching cost is actually higher though as G71 is MIMD in the VS. Vertex Texture Fetch is also likely to be more performant on Xenos because with G7x NVIDIA haven't built in any latency hiding to the texture lookup for VTF, which will stall the vertex pipeline; pixel shading is generally built with texture latency hiding mechanisms in place so on a unified shader vertex operations also benefit from this in operations that are likely to have latency bound instructions (i.e. VTF) and so it doesn't stall the pipeline.
mckmas8808
Legend
Dave Baumann said:
What about vertex shaders?
VT (Vertex Texturing) is just one possible way to create geometry. You can create complex geometry with oher methods such as Procedural content created on the CPU or even using R2VB to aleviate the VTF latency problems.predicate said:
It really can improve shadow mapping performances.expletive said:
predicate said:
RSX batch size not necessarily = G71 batch size. I thought I read somewhere that devs have some control over this on PS3, but I'm not sure.
I have a question, that's sort of on a tangent, but I didn't want to make a thread just for it - is texturing generally a consistent activity across a frame, or do texturing requests come in stops and starts? Also, G7x vertex shaders have their own (less powerful) texture samplers as opposed to contending with pixel shader texturing, yes?
Given that VS are generally “the less sexy†part of the pipeline, I’ve not really tried to follow them much. However, given that RSX is “NV47†based they are likely to be the same configuration as G71 and I would be surprised if the number of them changed.mckmas8808 said:
Branching in the shader is just like branching in any code: if “this†then do “thatâ€; so if a pixel has one property then you can carry on with one shader branch, if it has another then you can completely change what the shader is doing – i.e. you could change the lighting mechanism dependant on some surface properties that may be about, or change a shadowing sampling mechanism if you are at the edge of a boundary. If you using shaders that do these types of things then fast DB can bring performance benefits, although its not necessarily a requirement.expletive said:
The batch sizes are an controlled by the capabilities of the command processor, and my understanding is the command processor hasn't changed. This being the case, these type of properties are likely to be same between RSX and G7x.Titanio said:
Errm, Texture requests change according to dofferent commands and various parts of the scene will have different commands. Not sure whats being asked there.Titanio said:
Yes.
expletive said:
It'll likely be used largely as an optimization, there are a lot of newer algorithms that are expensive per pixel but that have an oportunity ot early out, so you can do less work for a lot of the pixels.
It's posibble it could improve quality aswell, rather than limiting some algorithm to a fixed number of iterations (hand unwound) for everything, you could use a tolerance to guarantee some quaility threshold.
Titanio said:
Texture demand is going to be based on shader/material selection, so it will vary accross a frame, it'll change perhaps 200 times a frame.
G71 certainly has an advantage in terms of available texture units, but the more complex the shader from an arithmetic standpoint, the less relevant the difference will be.
Bear in mind that the contention is potentially between batches on Xenos, not between vertex and pixel workloads, I would be extremely surprised if any batch doing vertex work was comparable in texture load to a batch doing pixel work. The worst case is almost certainly drawing extremly large Tris with multiple texture layers and minimal ALU Ops (say the futuremark multitexturing benchmark).
Xenos also has point sampling texture units available, that can be used for simpler vertex texture fetches.
Yes but the demo was made for the R520, which does pixels in batches of 16 IIRC.Hardknock said:
Offtopic:It kind of sucks that we have not and never will see a tech demo for the Xenos. Yes we have games but sacrifices have to be made when you're making a game especially early in the systems life.
