MS: "Xbox 360 More Powerful than PS3"

[liverkick]

Realistic physics are only going to help so much. It's the animation that needs a lot of work and that depends more on the artistic talents of the developer than any FLOP rating on the hardware.

What about procedural animation techniques, inverse kinematics, etc? Wont more advanced animation techniques require more processing to enable them? Endorphin's physics based Natural Motion could be an example of where these floating point heavy processors could help in the animation department.

 

[ROG27]

Is this really necessary?



It still doesnt change the fact that we havent seen anything on the playable on the PS3 that looks as good as GoW, regardless of why it looks pretty. I dont doubt that we will soon though.

I agree it would run smoothly on both consoles and i stated as much in my post, I think everything this gen will pretty much look the same on both consoles with some REALLY nice looking exlcusives on both sides.

Anyway, my response was mostly directed at your statement that releasing 6 months later should provide an order of magnitude more power and, for examples stated, i dont beleive its the case.

I never said anything about an order of magnitude more powerful...especially in the graphics department. In that area I think they will be pretty evenly matched (maybe with RSX having a slight edge). The differentiation will come in the form of CPU usability, which I believe, yes, CELL is an order of magnitude (or almost) more usable/powerful than the XCPU...just from the ridiculous things that it has demonstrated (on its own, unassisted many times I might add). If you haven't seen the medical imaging simulation CELL performs, you should--because Medical imaging is one of the most computationally intensive things you can do, outside of engineering-type math and simulation.

 

[a688]

I never said anything about an order of magnitude more powerful...especially in the graphics department. In that area I think they will be pretty evenly matched (maybe with RSX having a slight edge). The differentiation will come in the form of CPU usability, which I believe, yes, CELL is an order of magnitude (or almost) more usable/powerful than the XCPU...just from the ridiculous things that it has demonstrated (on its own, unassisted many times I might add). If you haven't seen the medical imaging simulation CELL performs, you should--because Medical imaging is one of the most computationally intensive things you can do, outside of engineering-type math and simulation.

Have you seen the demostrations in medical imaging using the xcpu? No? Me neither. You can't really compare them like that. Its like saying that a drag car is faster than say an F1 car. It is but only in a straight line.

 

[Titanio]

For physics, I think you're right, but it won't have a big visual impact.

Physics can have a HUGE visual impact, and we've already seen that.

For HDR, I'm pretty sure Xenos fully supports FP16 in addition to FP10, as that's what written in Dave's article.

His article does not say that. FP16 framebuffers are supported, but blending on them is not. FP10 will pretty much be the standard on X360.

On bandwidth, it really boils down to whether:

48GB/s - Framebuffer access - CPU usage is > or < or = to 22GB/s - CPU usage

That's going to vary from title to title, obviously. Will there be a typical case? Maybe. Unless someone can propose figures for the variables above, the discussion is kind of going nowhere - but even then, nothing that's proposed can be universally applicable.

 

[jvd]

Actually Cell can access the XDR ram seperate from link between RSX and itself. I assume RSX hangs off Flexio (as other theory makes sense) while the XDR is accessed through Cell's XDram controller.

It goes xdr - cell - flexio- rsx- gddr . So the fastest rsx can acess the xdr ram is the speed of wich xdr is attached to cell. Which is 20 something GB/s . Doesn't matter what the flexio runs at , your limited by the bandwidth from the ram. Which is going to be hit up by the cell chip already.

I don't think there's any reason to assume textures cannot come over from the XDR memory. If Cell can procedurally generate textures which most accept it can and send them directly over to RSX via the direct link they have I see no reason why textures residing in the XDR memory pool cannot flow over Cell's EIB just as procedurally generated textures would.

I never said it couldn't , it will only get a fraction of the bandwidth that the gddr pool gives. Because your limited by the bandwidth coming out of the xdr ram and then you have to subtract the bandwidth the cell is going to be using when acessing that memory. Then you might even have a speed reduction again in the flexio if your using that to transfer many procedurally generated textures and other post process effects .

Procedurally generated textures won't sent to RSX shouldn't eat into Cell's bandwith to the XDR but of course texture fetched from XDR would.

I look at bandwith for the GPUs as sort of a wash barring some new revelations. RSX has 40+Gb/s in aggregate bandwith while Xenos's bandwith it kinda tricky to nail down. Some of Xenos's 22.4Gb/s bandwith to main memory will be consumed by the needs of Xenon but then it has 32Gb/s to it's back buffer in the eDram which should perform AA and other bandwith intesnsive tasks. Xenon could consume up to 10+Gb/s of the bandwith to the main memory theoretically at least. It's hard to nail down...at least for me, but I think it's gonna be something of a wash as far as the GPUs are concerned.

Some of the rsx's bandwidth will be taken by the cell chip . Both the xdr bandwidth and flexio bandwidth will be used by cel land rsx when transfering any data the two need to share or give to each other before you get to texture bandwidth.

The question is how much bandwidth will cell consume

 

[dukmahsik]

maybe he was taking into account the dev tools making it a bit more powerful?

 

[ROG27]

Have you seen the demostrations in medical imaging using the xcpu? No? Me neither. You can't really compare them like that. Its like saying that a drag car is faster than say an F1 car. It is but only in a straight line.

There is a good reason that CELL is actually being used in the medical field. It is far more capable than what is currently available...inclusive of multi-core workstation type cpus. If MS or IBM had a competitive advantage with the Xcpu they would have leveraged it for this purpose. So implicity, we can infer that the CELL would produce superior results to the Xcpu without it actually running this application. For reasons unbeknownst to us, the Xcpu is not being used in the Medical field, or for anything other than MS's game console (probably because it's not as well suited to such applications).

 

[Fox5]

is that, regardless of what the hypethetical performance is, what Sony has shown thus far (real-time) seems to be far ahead of the curve of what Microsoft currently has to offer. This seemed to be flip-flopped last generation, as is to be expected. Simply enough, one of the best indicators of a competing platform's performance/power is the date which it hits the streets. Some might argue this logic...but let's be serious here. If something comes out 6 months to a year after something else, it's going to have a clear technological advantage.

Again, a console is a price/performance trade-off piece of machinery. Sony is opting for the higher-end this time around. Microsoft had to make trade-offs to get their system out the door on time.

They're both on 90nm though, so there's no guarantee that one will be significantly more advanced than the other since they have almost the same resources to work with.

 

[SynapticSignal]

That the PS3 could potentially be bandwidth limited in comparison to Xenos, since it has no EDRAM to offer more effective bandwidth, and even current gen games (like Splinter Cell or FarCry) have show large >30% drops when halving the bus speed on a G70.
.

you are right
and we can't take x speed from local rsx ram, then add y speed for xdr and take the results as overall bandwidh, because, the bandwidh critical tasks cannot be splitted in two different bus and memories (es blending, aa, hdr, and all the backbuffer operations)

so stop the "bus a + bus b = wow bandwidh" because in real world it can't help in BW bounded situations

 

[Titanio]

you are right
and we can't take x speed from local rsx ram, then add y speed for xdr and take the results as overall bandwidh, because, the bandwidh critical tasks cannot be splitted in two different bus and memories (es blending, aa, hdr, and all the backbuffer operations)

You can't split the framebuffer across two pools, at least easily*, but things like texturing? Yes. Effectively the max amount of bandwidth available to RSX across the two pools is additive.

*I guess it'd technically be possible to some work through one bus, then move the framebuffer to the other pool and work on it there, hence splitting the framebuffer work across both pool's bw. But you're not likely to need to do that..

 

[jvd]

You can't split the framebuffer across two pools, at least easily*, but things like texturing? Yes. Effectively the max amount of bandwidth available to RSX across the two pools is additive.

*I guess it'd technically be possible to some work through one bus, then move the framebuffer to the other pool and work on it there, hence splitting the framebuffer work across both pool's bw. But you're not likely to need to do that..

but its not that simple as you have a second chip (cell) acessing one or both of the pools of ram. Which reduces your bandwidth numbers

 

[Titanio]

but its not that simple as you have a second chip (cell) acessing one or both of the pools of ram. Which reduces your bandwidth numbers

Of course, that's why it's the absolute MAX amount RSX could take.

How much do you think the CPU would take?

10GB/s maybe? That'd leave RSX with 38GB/s. Assuming Xenon took the same amount on X360*, that'd leave 12GB/s for Xenos. Meaning that RSX's framebuffer could consume up to 26GB/s before it'd have less bw left over for "the rest" than Xenos. Applying the same CPU usage numbers across both systems, the more data the CPU eats, the more Xenos hurts in this comparison.

The problem with all of this is, there's no one set of figures that'll apply to every game.

*of course, bw consumption on the two chips is likely to be different. But assuming equally or more efficient bw usage (which may well be the case given the greater on-chip memory on Cell, and with data access being more explicit), Cell consuming more bw would simply imply it's processing more data (and hence more powerful), which some seem loathe to concede either.

 

[jvd]

IT depends on what the cell chip needs and is doing . Which is exactly why I said it all depends on how much ram the cell chip needs in the other post .

I could just as easily say the cell chip needs 18GB/s ram

That would make it look worse .

Or we can say 8GB/s ram and that would make it great. But we don't know . We don't know what the cell would be asked to do .

 

[Titanio]

IT depends on what the cell chip needs and is doing . Which is exactly why I said it all depends on how much ram the cell chip needs in the other post .

I could just as easily say the cell chip needs 18GB/s ram

That would make it look worse .

As long as you hold Xenon to the same standard, the figures work out pretty much the same in every case as above.

Saying you can't hold them to the same standard, that Cell will consume more, implies it is either

a) inefficient in terms of bw usage (i don't think this will be the case)
b) doing extra work that for example Xenos's eDram is doing (and isn't eating main mem bw on X360)
c) Cell is simply pumping more data in and out (is more powerful, at least if you assume the same processing, but on more data. Of course, it could still use any greater amount of processing power given the same amount of data as Xenon, but I'm looking at the more bw intensive alternative)

 

[jvd]

As long as you hold Xenon to the same standard, the figures work out pretty much the same in every case as above.

Saying you can't hold them to the same standard, that Cell will consume more, implies it is either

a) inefficient in terms of bw usage (i don't think this will be the case)
b) doing extra work that for example Xenos's eDram is doing (and isn't eating main mem bw on X360)
c) Cell is simply pumping more data in and out (is more powerful, at least if you assume the same processing, but on more data)

Your wrong . Xenon is limited to how much ram bandwidth it can use as its link to the xenos is slower than the link xenos has to ram .

So we can actualyl figure out how much of the main ram xenos will have . I believe its just over 10gbs and then there is the edram .

With the cell with just dont know as its not really limited by anything in how much ram it can use from either pool .

 

[Titanio]

Your wrong . Xenon is limited to how much ram bandwidth it can use as its link to the xenos is slower than the link xenos has to ram .

So we can actualyl figure out how much of the main ram xenos will have . I believe its just over 10gbs

It's 10.8GB/s both ways (21.6GB/s total, I think). Xenon could pretty much use all of the main memory's bandwidth if it wanted to.

 

[nAo]

For HDR, I'm pretty sure Xenos fully supports FP16 in addition to FP10, as that's what written in Dave's article.

Xenos doesn't 'fully' support FP16.

 

[jvd]

It's 10GB/s both ways. Xenon could pretty much use all of the main memory's bandwidth if it wanted to.

Are you sure about that , I'm pretty sure the bandwidth is lower going to the xenos from the cpu than it is from the ram .


Either way . IF we believe the hype of cell then a more powerfull cpu should need more bandwidth .

Of course if your like me and believe both chips are really close in performance than the bandwidth needs should be the same .

YOur pick

 

[Titanio]

Are you sure about that , I'm pretty sure the bandwidth is lower going to the xenos from the cpu than it is from the ram .

It's like 0.8GB/s lower:

 

[Joe DeFuria]

It's like 0.8GB/s lower:

There needs to be clarification, based on how the arrows are drawn in the diagram:

Is it simultaneous 10.8 GB/Sec to and from the CPU? (Dual single direction channels).

and

22.4 GB/Sec bidirectional from Parent to Ram? (1 bi-directional channel, than can max out at 22.4 GB/Sec in either direction)

If what I described above is correct, then while total bandwith the about the same, you can can have double the bandwidth in one direction from GPU to Ram...