What is PS4's 14+4 CU thing all about? *spawn

[function]

That's a bottleneck, not diminishing returns.

What do you think causes diminishing returns (none linear performance scaling)?

 

[mrcorbo]

That's a bottleneck, not diminishing returns.

Bottlenecks lead to diminishing returns. Workloads are variable as are their dependencies on system resources. As more and more of your workloads are being stalled by bottlenecks your performance will tail off until all workloads are effected. Hence, diminishing returns.

 

[Averagejoe]

It would probably be trivial to create a CPU-limited scenario where a GCN-based card with more CUs performed no better than one with less. If your performance is being held back by the lack of one resource, throwing more of another resource at it isn't going to add to your performance.

So jaguar is enough for 14CU but after that you get diminishing returns.?


But how many games are CPU bound.?

The whole balance argument doesn't even make sense,MS claim 14 CU is the right balance,but they have 12 CU not 14 CU working ones,sony try to stir developers to use compute the 14+4 was just and example and even Cerny say it.

 

[function]

He's just trying to demonstrate example of none linear performance scaling.

There are fixed function parts of the GPU that can, and do lead to an increased proportion of CU time being waste as you scale the number of CUs.

Edit: There are a couple of the developers here that have talked about situations where this can happen e.g. high vertex workloads

 

[mrcorbo]

Yeah that is why the 7790 performs better than the 7770,and the 7850 better than the 7790..

7970>7950>7870>7850>7790>7770

This is a line on PC each one with more or less CU than the other and in all scenarios the one with more CU performs better this is 100% accurate and we haves tons and tons of benchmarks proving this without shadow of a doubt.

Unless something is horrible wrong inside any of the next gen consoles this should hold pretty well.

This is what make the whole theory of diminishing returns so hard to swallow.

Two points:

1. There are other differences in those cards beyond the number of CUs.
2. Show me comparative benchmarks of those cards while in an 8-core-Jaguar-based PC... oh wait you can't. They don't exist.

 

[mrcorbo]

So jaguar is enough for 14CU but after that you get diminishing returns.?


But how many games are CPU bound.?

The whole balance argument doesn't even make sense,MS claim 14 CU is the right balance,but they have 12 CU not 14 CU working ones,sony try to stir developers to use compute the 14+4 was just and example and even Cerny say it.

You're confused. MS never said anything about 14 CUs being balanced. Sony did.

 

[Jay]

Yeah that is why the 7790 performs better than the 7770,and the 7850 better than the 7790..

7970>7950>7870>7850>7790>7770

This is a line on PC each one with more or less CU than the other and in all scenarios the one with more CU performs better this is 100% accurate and we haves tons and tons of benchmarks proving this without shadow of a doubt.

Unless something is horrible wrong inside any of the next gen consoles this should hold pretty well.

This is what make the whole theory of diminishing returns so hard to swallow.

Nothing has to be wrong, much less seriously.

Comparing PC parts is a nice useful start, but it's not the end result.
Was those cards benchmarked with powerful cpu's to make sure that only the gpu's was being tested, yes, and if not then that makes the cards comparisons to each other invalid much less consoles.
Just an example, not saying it's the cpu specifically that is under powered or anything.

A console is a built as a whole not individual parts.

The balance that Sony has gone for doesn't seem broken, wrong, or anything but well done to me.

Why would a diminishing returns on 14+4 be wrong if their vision of the future is to make good use of gpgpu?
It's not like the CU's are not going to not be doing anything, could even be doing graphic related stuff.

People really need to stop seeing things as broken, issues, etc.

Discussing what led to the design, and if you had x more bandwidth, or cpu, resource allocation etc, so there wasn't a diminishing returns for that specific use case is different than saying there is something broken or wrong with it.

 

[upnorthsox]

Bottlenecks lead to diminishing returns. Workloads are variable as are their dependencies on system resources. As more and more of your workloads are being stalled by bottlenecks your performance will tail off until all workloads are effected. Hence, diminishing returns.

Sure a bottleneck can lead to diminishing returns, but a diminishing return doesn't need a bottleneck to exist. They're different concepts. One (bootleneck) is a performance metric, the other (diminishing returns) is a value proposition. A value proposition is a numerical expression of an opinion, so that makes it even harder to define.

 

[function]

If you plot your performance on a graph vs CU's you'll see a none linear relationship. That is your diminishing return. That diminishing return is caused by increased bottlenecking.

This is the discussion we're having in this thread and the terms work perfectly well.

 

[mrcorbo]

Sure a bottleneck can lead to diminishing returns, but a diminishing return doesn't need a bottleneck to exist. They're different concepts. One (bootleneck) is a performance metric, the other (diminishing returns) is a value proposition. A value proposition is a numerical expression of an opinion, so that makes it even harder to define.

In this case the value is the work that can be done by engaging increasing numbers of ALU units while executing typical graphics workloads. There's no opinion there. There's a direct and measurable relationship between the amount of processing resources you are attempting to use and the work they are able to due within a given period of time.

 

[function]

Yeah that is why the 7790 performs better than the 7770,and the 7850 better than the 7790..

7970>7950>7870>7850>7790>7770

This is a line on PC each one with more or less CU than the other and in all scenarios the one with more CU performs better this is 100% accurate and we haves tons and tons of benchmarks proving this without shadow of a doubt.

Unless something is horrible wrong inside any of the next gen consoles this should hold pretty well.

This is what make the whole theory of diminishing returns so hard to swallow.

As you brought it up, lets take a look at 7970 vs 7870 in the PC space and see what twice the CUs and roughly twice the flops get you when triangle setup and rops stay the same.

http://www.anandtech.com/bench/product/548?vs=508

Obviously this is the PC space and there will be differences in the console space, but there's some big ass diminishing returns for you, seeing as you named the cards and talked about PC benchmarks ...

 

[upnorthsox]

Nothing has to be wrong, much less seriously.

Comparing PC parts is a nice useful start, but it's not the end result.
Was those cards benchmarked with powerful cpu's to make sure that only the gpu's was being tested, yes, and if not then that makes the cards comparisons to each other invalid much less consoles.
Just an example, not saying it's the cpu specifically that is under powered or anything.

A console is a built as a whole not individual parts.

The balance that Sony has gone for doesn't seem broken, wrong, or anything but well done to me.

Why would a diminishing returns on 14+4 be wrong if their vision of the future is to make good use of gpgpu?
It's not like the CU's are not going to not be doing anything, could even be doing graphic related stuff.

People really need to stop seeing things as broken, issues, etc.

Discussing what led to the design, and if you had x more bandwidth, or cpu, resource allocation etc, so there wasn't a diminishing returns for that specific use case is different than saying there is something broken or wrong with it.

How is 14+4 diminishing returns? What result do you believe is diminished?

 

[Airon]

I believe that a system can be defined "balanced" when each components do not act as recurring "bottleneck" in processing tasks.
Or that all the components are mixed in a way that is not limiting to the efficency and the performance of another component.

What is clear now, in term of balance, is that 14CU seems to be the perfect spot for graphic/rendering tasks.
We have 3 important clues in that regard.
1) Sony slides ("system balanced at 14)
2) Cerny interviews (system is not 100% rounded, extra ALU x computing et...)
3) ... Microsoft !

We are in fact forgeting that on the other part of the planet some MS guys seems to have reached the same conclusion: the system is somehow balanced at 12CU (or around there).
And to me, this is a great clue that must not be underestimated.

The point is.. Why 12-14 seems the perfect spot?
The 2 systems are similar and so it could be usefull to do some parallel reasoning about the 2 systems in order to understand what makes this 2 system balanced for graphic around 12-14CU.

I bet few coins on CPU...

It is also very interesting to know which would be the return in term of performance when all the 18CU are used for rendering. It will differ much from 14CU implementation?
Launch games, like KZ, use all the 18CU for rendering (I mean there is next to zero use of CU for compute in KZ x exemple).
PS4 is a super easy system and a marvel to work with, but nevertheless, in time, developers will learn to use it with more efficency.
In time they will also learn to use the CU for no-graphic tasks.
When in 2-3 year CUs will be used extensively for no-graphic tasks (as Cerny predicted) I suspect that games will not look worse than now due to the minor implementation of CUs for pure rendering tasks... It would be a pure no-sense.

This could be a good meter of the return in terms of graphic performance of the 18CU vs 14CU... Sony said "4 additional CU will have minor return if used for rendering" ... I suspect this is an euphemism.

 

[Jay]

How is 14+4 diminishing returns? What result do you believe is diminished?

See the reference to it in the Sony slide (assuming on first page of this thread).

I specifically said in a specific use case, I also said I think their system is well balanced as far as I'm concerned.

Therefore as I said, no diminishing returns if you are also doing gpgpu also.

 

[patsu]

Bottleneck is only one of the elements. Even without bottlenecks, we can still hit diminishing returns.

Diminishing returns is a concept in economics, along with opportunity cost, supply and demand.

If the game developer has some visual targets in mind, once those are more-or-less met, enhancing it further may not be as critical. It may take a lot more effort to achieve the last few percentage points.

What's more, developers can have more time during the next iteration/game for larger enhancements (e.g., different design, refactor code, new algorithm, etc,). Software is interesting in this sense.

They can spend the finite resources on other things that matters more to the users. By year 3-4, when the developers are more comfortable with the h/w, they supposedly can dive more deeply.

The 14+4 note is not so useful in the grand scheme of things. The developers can and will always find their own way to exploit all 18 CUs.

 

[east of eastside]

On top of the 14 + 4 CU split, does the PS4 have a GPU reserve for OS (from the 14) similar to Xbox (though likely less) putting the effective GPU gflops at 1.4 tflops or less?

Is this why PS4 launch games aren't doing the 60fps or 1080p they were expected to?

 

[patsu]

There is no 14+4 split in the GPU. ^_^
It'd be extra work for no/little benefits.

Developers may be advised to use the CUs wisely but there is no general formula here just because every game and project team are likely different.

The OS may reserve some resources under some circumstances, but no one knows either way yet.

 

[mrcorbo]

On top of the 14 + 4 CU split, does the PS4 have a GPU reserve for OS (from the 14) similar to Xbox (though likely less) putting the effective GPU gflops at 1.4 tflops or less?

Is this why PS4 launch games aren't doing the 60fps or 1080p they were expected to?

You're whole train of logic starts off wrong, so your conclusion is doomed from the start. There's no split.

How about this for spin? If the PS4 is "balanced" for using 14CUs for graphics rendering and has 18 total that means that is has more than it can effectively use. This would indicate to me (and someone correct me if I am wrong here) that when performing graphics operations PS4 should never be ALU limited. PS4's worst case performance is still significantly higher than what the XBOne is *theoretically* capable of if it were able to utilize 100% of the ALU capability of its 12 CUs.

 

[patsu]

The concept of balanced suggests some sort of equilibrium regardless of what the developers do or target. It's quite meaningless in real life. Developers will always work around h/w specs, focusing on their strengths. Their "ideal" system states may be different and much more dynamic.

 

[Shortbread]

On top of the 14 + 4 CU split, does the PS4 have a GPU reserve for OS (from the 14) similar to Xbox (though likely less) putting the effective GPU gflops at 1.4 tflops or less?

Is this why PS4 launch games aren't doing the 60fps or 1080p they were expected to?

The PS4 32 ROPs alone will insure a much smoother frame-rate amongst most 3rd party games, especially at 1080p. So regardless of how things are sliced & diced, the ROPs deficiency on XB1 can't be overcome by any magic math.

Removed bit leading to off-topic.