News & Rumors: Xbox One (codename Durango)

[XpiderMX]

He said differences on third party games will be 1 digit framerate (0-9FPS).

 

[AlphaWolf]

My understanding was that graphics tasks are so inherently parallel that throwing more cores at those problems tends to "Just Work". I'm sure it's not a perfect correspondence, but it's pretty dang close. Yes? No?

Yes and no. The limitations are a moving target. More CU's on graphics won't help if you're limited by bandwidth or cpu.

 

[3dilettante]

I don't think Albert P's recent post helped Microsoft's case much at all. While many of his points seemed valid, his first two seemed like pure hand-waving. Am I alone in thinking this? Am I misunderstanding how these things work?

Those first two points, heavily paraphrased, were:

1. 50% more CUs doesn't mean 50% more-bettah "graphics".
2. Boosting "each" CU by 6% is somehow better than boosting the entire GPU by that amount.

I think he's just saying that perfect scaling based on core count doesn't happen, and that when counting CUs people are missing that they aren't counting the exact same thing because of different parameters for the designs.

 

[bkilian]

Those first two points, heavily paraphrased, were:

1. 50% more CUs doesn't mean 50% more-bettah "graphics".
2. Boosting "each" CU by 6% is somehow better than boosting the entire GPU by that amount.

My understanding was that graphics tasks are so inherently parallel that throwing more cores at those problems tends to "Just Work". I'm sure it's not a perfect correspondence, but it's pretty dang close. Yes? No?

I don't understand what he's trying to imply in his second point, at all. I assume he's just confused, but I'd like to hear if what he's saying makes any more sense to the experts here.

The two points are related. When you add cores, you always add some level of overhead to handle the extra core. Two cores can't do twice the work of one core, even in a highly parallel problem, since the are always places where a bottleneck occurs. Let's say each added core makes all your cores 99% as efficient as they were before you added the new core. At 12 cores, you're looking at each core being about 88% as efficient as it would be by itself. At 18 cores each core would be 83% of its best. This is the same reasoning behind why beefing up all your cores could be better than adding a new core.

I don't actually know what the real efficiency loss is per core added, but the principle is the same, irrespective of how small the difference between "perfect correspondence" and actual performance is.

 

[mrcorbo]

To the hardcore non family gamer yes your right .
But to the family man who started playing games in the 70s and 80 now has two point two children of different ages . A wife who enjoys social media and small games of her tablet .its not such a hard sell as everyone in his family can enjoy what the Xbox one offers .

Like I said Microsoft are looking to a different market the family that's why the hardcore gamer is spitting his dummy out on the internet the biggest company in gaming is saying your only part of our vision .where as there rival is saying we still love you .
Again its just my opinion on the available facts as I see them .

And I don't see this product appealing to that market. Not at that price point.

 

[mrcorbo]

The two points are related. When you add cores, you always add some level of overhead to handle the extra core. Two cores can't do twice the work of one core, even in a highly parallel problem, since the are always places where a bottleneck occurs. Let's say each added core makes all your cores 99% as efficient as they were before you added the new core. At 12 cores, you're looking at each core being about 88% as efficient as it would be by itself. At 18 cores each core would be 83% of its best. This is the same reasoning behind why beefing up all your cores could be better than adding a new core.

I don't actually know what the real efficiency loss is per core added, but the principle is the same, irrespective of how small the difference between "perfect correspondence" and actual performance is.

A question occurred to me along these lines when I saw the original quote.

In the context of a GPU performing typical graphics operations, is it reasonable to argue that a % increase in theoretical maximum performance by increasing clockspeed is not directly comparable to a % increase due to increasing functional units?

 

[3dilettante]

Neither has ideal scaling, although a clock bump does have the additional benefit of scaling up various resources that wouldn't scale if only CU count was increased. Durango's GPU front end, eSRAM bandwidth, and internal bandwidth went up as well.

There are still things that may not have changed, like the off-die bandwidth and the links to the rest of the chip. Higher clocks would have a higher fraction of idealized performance lost due to the portion of the system that didn't improve.
However, the percentage change is a small increment, so it's easier to hand-wave a rounding error with a single-digit percentage than it would be for a 50% unit increase.

That does mean that the benefit of said incremental increase should be viewed for what it is: something nice to have that doesn't change the overall picture much.

 

[Rangers]

Lastly, and changing the subject, how crazy would it be if Microsoft is being prevented (via an NDA) from talking about some super-neato aspect of their hardware? Wouldn't that just perfectly fit in with the rest of this "rolling thunder" of a PR disaster they've been executing on? "We've released all the unimpressive parts of the specs, but are prevented from discussing the good parts. Please understand."

I'm not saying I believe that's what's happening, by God, wouldn't that just be par for the course this year?

Well, I've heard two rumors, since this is the rumors thread.

As each of the larger rumors such as 12GB RAM, dGPU, get killed off, maybe the rumors get smaller

But anyway one is the XGPU has enhancements from AMD"s upcoming Volcanic Islands GPU tech. This is said to be a minor help not a gamechanger.

And another is there is an extra 2 CU's above expected in X1? I'm not sure the reasoning about this one since 1.31 TF was clearly on display at hotchips. But I guess the theory would be of enabling redundant CU's.

He said differences on third party games will be 1 digit framerate (0-9FPS).

Penello actually said that? Could you provide a link?

That's actually not hugely positive as in a typical 30 FPS console game 5 FPS is a large percent of the total (well, 16.6% in that case). It would also be the first explicit admission by Microsoft, of a weaker system.

 

[warb]

That'd be less than it was with some PS360 games.

 

[BRiT]

Of course he never said that they would be running identical resolution or settings. If you target a certain FPS then of course there won't be much difference between the systems.

 

[Rangers]

That'd be less than it was with some PS360 games.

Well, 0-9 FPS IF talking about 30 FPS is 0-30% for what that's worth. I guess even the high end is under the 40% GPU FLOPS deficit.

 

[warb]

Well, 0-9 FPS IF talking about 30 FPS is 0-30% for what that's worth. I guess even the high end is under the 40% GPU FLOPS deficit.

We've seen ~5fps plus fairly significant image quality differences in more than a few games over this generation.

 

[Rangers]

OK I apparently found the post in question for better clarity.

I will ask two questions of the detractors, honest questions.

1. What piece of information would you want that I could provide that would convince you there is not a huge delta in performance?
2. If it comes out after we launch that the difference between 3rd party games is maybe single-digit FPS between the two platforms, will I get an apology or concession?

 

[Strange]

Well, 0-9 FPS IF talking about 30 FPS is 0-30% for what that's worth. I guess even the high end is under the 40% GPU FLOPS deficit.

30 fps framelocked and 30 fps vsync with frequent dips into 21fps regions is a pretty big performance difference.

He really shouldn't have suggested the single digit fps difference.

1. What piece of information would you want that I could provide that would convince you there is not a huge delta in performance?
2. If it comes out after we launch that the difference between 3rd party games is maybe single-digit FPS between the two platforms, will I get an apology or concession?

He really should shut up.
Do consumers get an apology or concession if the difference between 3rd party games is greater than single digits?
(like 60fps framelocked versus 49 fps, which is 18.3% difference in fps, and could very well be possible be the case with 30~40% GPU difference)

 

[Jay]

This seems to have links to his posts regarding it all:
http://www.reddit.com/r/xboxone/comments/1m2suv/penello_specifies_why_x1vsps4_performance/cc595y6

He really shouldn't have suggested the single digit fps difference.

Agree with you there, especially as a single digit has a pretty big range.
One of those comments that probably sounded better in your head

 

[Strange]

Agree with you there, especially as a single digit has a pretty big range.
One of those comments that probably sounded better in your head

Not only that, but going deeply it's also a confession of inferior performance.
Which he shouldn't be putting himself in the position to confess in the first place.

 

[ERP]

FWIW my expectation is that PS4 ought to have a performance advantage, but I wouldn't expect it to reflect the difference in CU counts. CU's are MASSIVELY underutilized on vertex heavy workloads and plenty of the frame will be ROP or bandwidth limited.
There are just too many little things that can significantly impact performance, there were times with early firmwares on PS4 where seemingly innocuous changes would affect performance by as much as 10%.

The eSRAM will certainly provide an advantage under some circumstances, and I'm interested if the ROP difference will end up being a factor, or the lower eSRAM latency will end up nullifying it.

For that matter I could imagine the graphics API having a significant effect, the CPU single threaded performance isn't great on these machines and a poorly conceived implementation, could hurt games across the board. Sony having "lower level access" here isn't necessarily a win.

Hell I could imagine cases where CPU limited games demonstrate an advantage on XB1.

 

[Silent_Buddha]

Well, 0-9 FPS IF talking about 30 FPS is 0-30% for what that's worth. I guess even the high end is under the 40% GPU FLOPS deficit.

Are there any launch day multiplatform games targeting less than 60 FPS at the moment? The only ones I'm aware of that have mentioned FPS are all targeting 60 FPS.

Regards,
SB

 

[XpiderMX]

Not only that, but going deeply it's also a confession of inferior performance.
Which he shouldn't be putting himself in the position to confess in the first place.

Penello never said X1 isn't inferior, just not 30%-50% as forum (gafers) said.



I guess 1~3 frames of difference is not a 30% weaker like forums said.

 

[McHuj]

FWIW my expectation is that PS4 ought to have a performance advantage, but I wouldn't expect it to reflect the difference in CU counts. CU's are MASSIVELY underutilized on vertex heavy workloads and plenty of the frame will be ROP or bandwidth limited.

Out of curiosity, and I'm not a graphics programmer, is it not possible to pipeline the graphics work such that vertex processing for frame "n+1" is happening while ROP's are processing pixels for frame "n"?

Are there in effect two sets of "efficiencies": 1. Of the time running shader code, how much of the shader's processing is utilized using useful work. and 2. Over the course of frame, how often is each unit (shader/rop/etc) active/inactive.