MS: "Xbox 360 More Powerful than PS3"

[Titanio]

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...

Yes, but Xenon can read 10.8GB/s and write 10.8GB/s, which the GPU->RAM bw would accomodate, and leave 0.8GB/s left over. Xenon couldn't saturate the bw going one way, but it could between a combination of its max read/write bw.

 

[scooby_dooby]

So I suppose - applying your logic - PS2 is hands-down the most powerful platform because it yields a bandwidth advantage of what? 20 to 1?

When did I say X360 was hands down the most powerful platform? When? When did I say BW was the defining factor in a systems power? Do you just make this stuff up?

Either backup what you say, or quit accusing me of stuff I didn't say. That's pretty simple.

I said it seems to me Xenos has hands down a BW advantage,(due to the fact with it can make use of 256gb/s bandwidth within edram) I even added that qualifier that it would only be when tiling was implemented correctly and efficienty.

I don't see the problem with that statement, and if you think it's wrong, by all means explain your POV.

 

[Gholbine]

Guess that settles it then!

 

[scooby_dooby]

Silly boy, GoW would run smoothly on both platforms.

Ok so let me get this straight. When Sony shows tech demo's your comment is:
"what Sony has shown thus far (real-time) seems to be far ahead of the curve of what Microsoft currently has to offer."

But when MS shows a jaw dropping REALTIME GOW cut scene you say:
"Silly boy, GoW would run smoothly on both platforms"

Do you see the hypocrisy? Don't you see how you're a "silly boy" because all the sony "realtime" demo's could also be done on the 360!?

Or do you think that realtime engine's dev's have created 6 months before launch for PS3 are already superior to what will be done over the entire lifespan of the 360?

You make absolutely no sense. Everything we've seen thus far could probably be done on either system.

 

[j^aws]

It's like 0.8GB/s lower:

The above diagram is misleading as it doesn't show any blocking occurring with bus accesses. This was discussed in the thread below. The conclusion was that peak access to GDDR at 22.4 GB/s blocks CPU-Parent_GPU to 10.8 GB/s (not 10.8+10.8)

i.e. 10.8 GB/s from L2 cache + 22.4 GB/s from GDDR is 'peak' (33.2 GB/s aggregate). Just look at the leak diagram form 2004 (bottom right corner '7'), and the more detailed data flow diagram below,

http://www.beyond3d.com/forum/showthread.php?t=23775

 

[dubyateeeff]

"On balance I think most people who study the space will tell you our system is slightly more powerful then theirs from a hardware standpoint..""
'
That quote is cleverly worded. Basically he's claiming the 360 is more balanced not more powerful.

But its true, it is definetly much more powerful than my system, I have a Athlon 64 with a R9800 256MB Puro. What about you, is your system more powerful?

 

[ihamoitc2005]

Internal bandwidth

I said it seems to me Xenos has hands down a BW advantage,(due to the fact with it can make use of 256gb/s bandwidth within edram) I even added that qualifier that it would only be when tiling was implemented correctly and efficienty.

If tiling is implemented main memory bandwidth is consumed regardless of efficiency of such method. Therefore Xenos has hands down bandwidth disadvantage for rendering games at HD resolution.

 

[Laa-Yosh]

If tiling is implemented main memory bandwidth is consumed regardless of efficiency of such method. Therefore Xenos has hands down bandwidth disadvantage for rendering games at HD resolution.

You do know that tiling will only take a few hundred MB/s from the total bandwith at worst, right?

 

[Bill]

Once you look past the PR bullshit, you'll recognise that different sets of hardware pose different trade-offs, different philosophies and different strengths. Does that answer your question?

And then you'll realize one is still more powerful than the other.

 

[Fafalada]

You do know that tiling will only take a few hundred MB/s from the total bandwith at worst, right?

No.
At worst, tiling would take 'number of tiles' times more bandwith then a single tile.

But one reason why we have smart people writting games is to get as close as possible to best cases for utilization, rather then the worst ones.

 

[Bill]

My take is that both Sony and MS are taking their stances on the concept that whoever releases their console last typically has the more powerful technology.

Time really had nothing to do with it.

MS is using 337 million transistors on the 90nm GPU developed about the same time Sony is using 300m. I think six months bought Sony a small clockspeed bump maybe..that's about it.

If they're significantly oupowered, it's down to bad engineering. This would be the EDRAM I'm always griping about in my opinion. While MS is using 337m, only 232 is usable shader logic, a significant deficit.

 

[TurnDragoZeroV2G]

While MS is using 337m, only 232 is usable shader logic, a significant deficit.

Except that far less than that is "shader" logic. There's plenty of other things on the main die. Memory controllers, texture pipes, cache and registers, etc. I think the logic that actually goes into the programmable shader execution logic is a considerable fraction, but by no means 80% of it or anything.

But as long as the 232million all counts as something, then why not the 20-25 million worth of logic on the daughter die. Besides, that logic constitues just half as many ROPs, with fixed-pattern sampling and greater simplicity versus r520, g70, etc. (though they do support up to four samples/clock, IIRC) I don't think it'd be unfair to say that one would end up with 272 million if there were no eDRAM, and that's just from doubling up ROPS. Given it's also assuming that trading a programmable pattern, fp16 blending, etc. for four samples a clock resulted in transistor savings or an equal trade.

 

[Bill]

The biggest question about RSX is the split BW. Is it additive or in effect confined to 22 GB/s?

Nobody here seems to have answered this.

If what Titanio is saying, more like additive. They could do textures from XDR. Then shaders in GDDR.

Anyway, whenever I bring up EDRAM people say it was necessary. The alternative would have been far too costly. But I see no financial reason MS couldn't have gone with the same split setup as Sony. (neither company wanted to use 256 bit bus). In fact I recall MS mentioning they considered it, but didn't want to force developers into using 256 for this and 256 for that, but rather preffered to give them 512 however they wanted. But I suspect the EDRAM was a done deal, so it wouldn't have mattered.

But as long as the 232million all counts as something, then why not the 20-25 million worth of logic on the daughter die. Besides, that logic constitues just half as many ROPs, with fixed-pattern sampling and greater simplicity versus r520, g70, etc. (though they do support up to four samples/clock, IIRC) I don't think it'd be unfair to say that one would end up with 272 million if there were no eDRAM, and that's just from doubling up ROPS. Given it's also assuming that trading a programmable pattern, fp16 blending, etc. for four samples a clock resulted in transistor savings or an equal trade

The problem is some of that is fixed function AA. I'm not sure laypeople will appreciate AA that much. It's also not flexible. RSX can dedicate power to AA or power elsewhere.

I have no idea how much ROPS use, but I figured about 10m per set of 8. I figure 10m of the daughter die would have been needed anyway. So 242.

Regadless I would have rather had that EDRAM used on more execution resources.

Unless X360 games end up looking as good or better than PS3. There's no excuse not too.

I have yet to see a good console design that uses EDRAM. So far it's been contained in GC, PS2, and X360. So far none have impressed me to put it mildly.

 

[Farid]

If they're significantly oupowered, it's down to bad engineering. This would be the EDRAM I'm always griping about in my opinion. While MS is using 337m, only 232 is usable shader logic, a significant deficit.

I stop counting the number of times you said that, each time people explained that eDRAM was an architecture choice, and that it was not due to "bad engineering". All architectural choices have pluses and minus, and that's it.

Now enough with that.

 

[Bill]

Anyways what about the latest rumors RSX is 32 pipes?

INQ is saying G71 is 32 pipes now. They've said in the past it's RSX.

Or I could see a disabled qaud..28 pipes..in any case MS would be screwed. Unless the BW does hurt Sony a lot..

I stop counting the number of times you said that, each time people explained that eDRAM was an architecture choice, and that it was not due to "bad engineering". All architectural choices have pluses and minus, and that's it.

Then it can be a BAD structural choice.

Again, 337m transistors on 90nm. 300 on 90nm.

X360 games better look as good or close to it. Or something failed.

And all breakdowns post the GPU costs as roughly the same as well in the two consoles.

 

[zidane1strife]

Hmmm now that were in the talk about bandwith, comparing consoles and what not, I thought I'd bring up this comment:

So the final product will be an 1080p game? Running at a high target FPS?

Will it use HDR?

Sony asked for 1080p for the trailer, so we provided it... Assuming all things being equal I can't see why the final game wouldn't...

FPS I'll be really disappointed if we can't keep a steady 30 at worse and would prefer 60 but given time scales 60 may not be possible (I'm personally fussy about framerate so I'll try for a steady 60fps)

HDR is bread and butter for us, something like 2 years old tech. Originally 16 bit integer (back in the ATI 9800 days) now FP16 based. X360 has a nice 32 bit float mode (FP10) but I expect that will be quite hard to use for proper HDR (never got a chance to use it so can't really comment), considering we have actually hit precision issues with FP16 framebuffers (64 bit). FP16 seems to be the best comprimise for a framebuffer at the moment (its enough for most lighting, though we run out of precision directly in front of the sun (noticeable on our cloud renderer))

Ok, if I'm getting this right a highly demanding title with lots of stuff going on and extremely impressive looking, with iirc heavy particle effects, motion blur, depth of field, hdr at 1080p running at 60fps is not entirely out of the question? That is it is not entirely dismissed from the get-go due to technical qualifications?

 

[Mintmaster]

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

I never said it wouldn't, but we're not talking about no physics vs. awesome physics. My point is 2 or 3 times the FP power (doubtful) will not have a significant impact on physics, especially in terms of what it brings you in a game.

For example, the LU factorization IBM showed is a n^3 algorithm. If you do brute force interaction, 2x the FP power buys you only 26% more objects. The algorithms that a physics engine uses make a huge difference, and these algorithms will be 10 times harder to optimize for Cell. You don't have a predictable set of data to stream in.

I've actually coded a physics engine. Yes, there's a lot of FP code in there, but really impressive physics (which I haven't achieved yet) needs to navigate and maintain large spatial data structures, solve sparse matrices, and know how to leave things alone that aren't going to change. It's not simply a matter of crunching numbers.

 

[jvd]

Hmmm now that were in the talk about bandwith, comparing consoles and what not, I thought I'd bring up this comment:




Ok, if I'm getting this right a highly demanding title with lots of stuff going on and extremely impressive looking, with iirc heavy particle effects, motion blur, depth of field, hdr at 1080p running at 60fps is not entirely out of the question? That is it is not entirely dismissed from the get-go due to technical qualifications?

The video shown was not running at 30fps + as was shown. It was sped up and he admited it .


Although he has said that the game is at playable framerates now , we haven't seen anyone playing it (or any new shots actualy ) to prove it . This is not a knock against deanoc as I respect him. But we haven't seen anyhting so far

 

[Fafalada]

solve sparse matrices

IMO* Cell should spank a conventional CPU silly when it comes to matrix solvers.

*While I have not written an entire Physics engine by myself (frankly I don't think anyone in this day and age can really do that if you want reasonably featured thing produced in a reasonable time) I've spent enough time on lower level stuff like matrix solvers that I have some idea how they tend to be affected by underlying CPU architecture.

That said, my reasoning for above comment isn't limited to FP spec.

 

[Mintmaster]

Bill, you really are full of misinformation.

If they're significantly oupowered, it's down to bad engineering. This would be the EDRAM I'm always griping about in my opinion. While MS is using 337m, only 232 is usable shader logic, a significant deficit.

Because they used eDRAM, they don't need the memory controller to be as fancy, they don't need colour compression or z-compression, they get peak fillrate regardless of whether alphablending is used or if AA is enabled, and they save wiring cost. They also get higher yeilds from two smaller chips. Current speculation is that there's redundancy in there too, so they really crammed 64 shader pipes into those transistors.

The disadvantages are non-unified memory (more copying) and packaging cost. Xenos also took the responsibility of the memory controller from Xenon, so don't forget to factor that in as well.

The problem is some of that is fixed function AA. I'm not sure laypeople will appreciate AA that much. It's also not flexible. RSX can dedicate power to AA or power elsewhere.

AA is a fixed function feature on both RSX and Xenos. There's no way to allocate that power elsewhere. RSX needs colour compression as well for AA to run fast, and that takes up a fair amount of logic. The only power that can be freed by not doing AA is bandwidth.

The biggest question about RSX is the split BW. Is it additive or in effect confined to 22 GB/s?

Nobody here seems to have answered this.

If what Titanio is saying, more like additive. They could do textures from XDR. Then shaders in GDDR.

Do you know where pixel shaders get their data from? The answer is... Textures! Your proposition makes no sense.

Graphics pipelines have a lot of transistors devoted to holding pixels in flight in FIFO's. A texture access takes a long time because the memory controller waits for a group of accesses and puts it all together. While the shader pipe is waiting, it's working on another pixel, but obviously there's a lot of data to store about that pixel. RSX has direct access to GDDR3, and that's what it'll be using 99% of the time. It would be silly to waste transistors to have good performance with memory access from the XDR. Vertex data can come from there no problem, but texturing will in all likelihood impose a performance penalty.

There is not going to be a blowout this generation from either side. Perception wise, however, Sony has it in the bag. I'm sure half the population thinks PS3 will be at least twice as powerful as XB360.

In the end it'll come down to developers.