PlayStation 4 (codename Orbis) technical hardware investigation (news and rumours)

[Betanumerical]

I'd argue that it was commonly conflated as suggestive of pure graphics power* from day 1 when that isn't necessarily appropriate and that this caused the issue (similar to PS4's OS reserve for RAM).

*Specifically, 'graphics power' in the sense of visual output on screen and tech graphics performance. In other words, it was being proclaimed all over the place that PS4's 18 CU's would lead to '50% better performance' than the 12 CU's in X1.

So in that sense I'd say it was a positive improvement to PS4 from a design pov, but the PR and common rhetoric surrounding the issue suddenly raises the specter of an apparent downgrade once ppl discover that not all of those 1.84Tflops will be used for actual graphics.

This is absolute rubbish once again, how many times does it need to be said that you can use all 18 CU's for graphics? and either way, even if you aren't going through the graphics pipelines compute jobs can still be used for graphical effects...

I could just as easily say that not all the 12CU's on the XBONE could be used for actual graphics but im sure youd have some BS reason why it doesn't apply to your preferred console.

 

[Boxercide]

Yeah, no agendas here...

Tell me more about the 14+4 split and how this proves your POV.

 

[temesgen]

The whole discuss is basically rubbish, Sony given their financial situation isn't likely to be inclined to waste money on CUs that won't be utilized. The PS4 has more CUs and ROPs but for some reason we're having a debate about the ability to use that power because someone strings some sentences together. The engineers at Sony and MS must be idiots if what we're discussing is accurate.

In the case of MS, someone thought they were designing a chip to do one thing but magically it is able to do more (sailboat becomes a submarine) and in the case of Sony the engineers put their decimal in the wrong place and cost the company millions due to over engineering a dangerously out of balanced machine (oopsy.... ).

The facts are engineers work with very specific constraints and goals set forth by the overall intent which drives the vision which serves as the basis for the design. Both groups of engineers where handed budgets and expectations for performance and then created solutions with all these factors involved.

 

[MJP]

It doesn't really make sense to couple CPU throughput with GPU throughput. In general the CPU will feed the GPU commands, so you could possibly make the argument that your CPU throughput puts a cap on the number of commands you generate for the GPU in a particular frame. However there's not really a direct relationship between the number of commands issued and the amount of work performed by the GPU. You can very easily set up a single Draw or Dispatch call with a complex shader that will run on thousands or even millions of threads, and that one call will completely saturate the GPU for several milliseconds.

Even if that weren't true, the overhead from generating commands is typically much lower on consoles than it is on PC's. Consoles present lightweight API's specifically tailored to the hardware, and often ofter direct memory and command buffer access. On PC you have to go through heavy drivers that do lots of work behind the scenes to make sure the GPU behaves according to the model presented by the graphics API. Combine this with the fact that on consoles you can typically pre-generate command buffers if you want, and you pretty much have to throw any direct PC comparisons out the window. However even on PC you have options for reducing draw calls and state changes, by using things like instancing and texture arrays or even compute shaders that batch up geometry into a single draw call.

 

[Rangers]

The whole discuss is basically rubbish, Sony given their financial situation isn't likely to be inclined to waste money on CUs that won't be utilized. The PS4 has more CUs and ROPs but for some reason we're having a debate about the ability to use that power because someone strings some sentences together. The engineers at Sony and MS must be idiots if what we're discussing is accurate.

In the case of MS, someone thought they were designing a chip to do one thing but magically it is able to do more (sailboat becomes a submarine) and in the case of Sony the engineers put their decimal in the wrong place and cost the company millions due to over engineering a dangerously out of balanced machine (oopsy.... ).

The facts are engineers work with very specific constraints and goals set forth by the overall intent which drives the vision which serves as the basis for the design. Both groups of engineers where handed budgets and expectations for performance and then created solutions with all these factors involved.

First of all Cerny's comments really are kinda obvious about this. So it's not really too reasonable to debate.

Digital Foundry: Going back to GPU compute for a moment, I wouldn't call it a rumour - it was more than that. There was a recommendation - a suggestion? - for 14 cores [GPU compute units] allocated to visuals and four to GPU compute...

Mark Cerny: That comes from a leak and is not any form of formal evangelisation. The point is the hardware is intentionally not 100 per cent round. It has a little bit more ALU in it than it would if you were thinking strictly about graphics. As a result of that you have an opportunity, you could say an incentivisation, to use that ALU for GPGPU.

In the context of our current discussion, Cerny's comments fit perfectly.

Second, nobody ever said the +4 CU's were "unusable" worst case scenario, except you just then, because apparently everybody thinks graphics are the only thing that matters.

 

[Betanumerical]

First of all Cerny's comments really are kinda obvious about this. So it's not really too reasonable to debate.



In the context of our current discussion, Cerny's comments fit perfectly.

Second, nobody ever said the +4 CU's were "unusable" worst case scenario, except you just then, because apparently everybody thinks graphics are the only thing that matters.

Except it makes no sense because you could easily use GPGPU/compute for graphical effects, what extra juice exists in the other 4 CU's for compute I can honestly not think of, maybe its because compute allows you to do more, or to do things differently but they are the only two reasons I can think of.

Also if we really want to base everything we are saying strictly off Cerny's comments does that mean that the PS4 GPU is 1.84TFLOPs for graphics and 1.84TFLOPS+x for compute? because thats what he is saying, and it sounds a little ridiculous tbh.

 

[pMax]

As said already, some tasks tend to get by with a low amount of CUs (like shadow map rendering iirc). When this is done, one could allot more CUs for compute tasks, but during the post processing stage (which one can perfectly distribute over as many CUs as there possibly can be) all CUs are devoted to this and no compute gets scheduled at all for example.

This is quite more interesting than the +4 BS discussion, imho.

 

[borntosoul]

Wow, maybe it means that in some cases the CPU can get a bit of help ,

eg; in an environment with 1080p limit some CUs may be better used to do something else.

Does that explain it?

 

[pMax]

CUs are not processor cores, and thinking of them like processor cores is plainly wrong.
You should think of them as 'available resource in pool', with a shared front end that manage them all at once.

So, referring to CUs as 'units' looks very strange to me - they are not 'individual cores' where you send stuff to execute and so on, like you set thread affinity and such, even if they can 'run' a sequence of instructions by their own.

To make a comparison, is like if you were referring not to the processor cores, but to one of the ALU of a processor - would it make any meaning to you??

 

[3dilettante]

CUs are not processor cores, and thinking of them like processor cores is plainly wrong.
You should think of them as 'available resource in pool', with a shared front end that manage them all at once.

The distinction isn't as stark these days. The CUs aren't fully featured cores, but they have a mix of features that make them appear like hybrids of more advanced shared memory processors at the high end and very simple embedded controllers at the low end.

They are multithreaded, work in a shared virtual address space, and have a read/write cache hierarchy, all of which places them higher than the simplest cores.
They don't really service interrupts or have much in the way of exception handling and context switching, and their heavier reliance on a separate hardware unit to initialize their kernels keeps them distinct from the higher-end generalist cores.

So, referring to CUs as 'units' looks very strange to me - they are not 'individual cores' where you send stuff to execute and so on, like you set thread affinity and such, even if they can 'run' a sequence of instructions by their own.

Thread affinity is, for now, implied because threads don't migrate. The low-level access provided by the consoles and the PS4's customizations in general may allow for more direct work allocation to specific CUs. There may eventually be some extensions to the compute APIs to provide a bit more control as well, although I haven't kept up with that in great detail lately.
Even without it, given the large number of RISC cores that exist in the average PC, it's not a hard requirement that the software has to see a core's instruction stream to make it one.

Given the wide-ranging taxonomy of what makes a core. Running its own instructions is a very big chunk of what would make a piece of silicon a core.

 

[temesgen]

First of all Cerny's comments really are kinda obvious about this. So it's not really too reasonable to debate.



In the context of our current discussion, Cerny's comments fit perfectly.

Second, nobody ever said the +4 CU's were "unusable" worst case scenario, except you just then, because apparently everybody thinks graphics are the only thing that matters.

There have been plenty of allegations that 4 of the CUs couldn't be used for graphics, perhaps the person who hammered that point the most was Xenio. BTW Xenio apparently drew those conclusions based off what he read in the leaks if his post are to be taken literally, we see similar logic being applied to PS4's audio capabilities...

To your point about not using of the CUs for graphics again why would the engineers put so many CUs and ROPs in the machine? The overall design of PS4 suggest it has a higher ceiling where visuals are concerned. Does it make more sense to over engineer the GPU and force it to handle audio or throw in a DSP? Which is cheaper? As many have pointed out the use of the CUs will be a function of the goals of the developers - some might use them all for graphics while others might use some for physics for example.

Finally you said earlier in the thread that XB1 has 40% more CPU - walk me thru the math on that. I've seen that said before and I've also seen it disputed so please reply with more of detail so I can see how you arrive at that conclusion. XB1 may very well have more CPU power in the box but I am curious why MS would spend money there if they think cloud computing can truly have the impact they are suggesting for compute. Why not use a cheaper part and save on the BOM?

The glass is half full interpretation where my platform of choice is concerned and glass half empty interpretation of the competitor stuff is getting a bit stale. Its somewhat disingenuous to assume the absence of information must mean the worse case scenario for one platform while the same lack of clarity is being use to suggest your platform of choice could have more power that hasn't been disclosed. Maybe we should submit this as a new principle in computing, we could call it Fanboys Law? There is a small vocal minority in these threads who do just that in very predictable ways, repeatedly over and over and as someone mentioned earlier its sad.

 

[pjbliverpool]

I agreed with expletive's comment on this argument earlier. Resolution is VERY often associated directly with diminished returns. Devs know this even though many PC gamers refuse to accept it as true, let alone obvious. Hence so many console games ship with resolutions slightly lower than 720p (devs realize it's not worth processing all those pixels on a limited computing budget when ya can't easily see the damn things).

Maybe Sony/AMD felt that throwing more pixels on screen beyond what is typically doable in a modern game setting at 1080p just isn't worth it. Wouldn't surprise me, after all, that is the reason MS made those display planes in X1. Likewise, I'd assume at some point a certain level of AA/AF becomes basically not worth the processing effort. I'd imagine those are the things AMD/Sony may have been thinking of when considering how to balance the design.

Surely you're not suggesting that the only use for more than 14 CU's is going to be increased resolution and AA/AF?

so it's no wonder many still hold that short sighted view that the real world comparisons to be drawn should rely on a GPU with 1.84Tflops dedicated to graphics rendering as a baseline feature. To someone coming from that pov it would seem like a downgrade, even though it isn't.

It's not a "short sighted view", it's cold hard fact. PS4 has 1.84 TFLOPs which can be dedicated to graphics or some mix of graphics and GPGPU just as the XB1 has 1.31 TFLOPs which can be dedicated to graphics or some mix of graphics and GPGPU. This is a completely fair and accurate comparison. To suggest otherwise is not "taboo" on B3D, it's simply wrong. And people spreading false information over and over in spite of having been corrected by numerous people on many occasions is what's frowned upon here.

I'd argue that it was commonly conflated as suggestive of pure graphics power* from day 1 when that isn't necessarily appropriate and that this caused the issue (similar to PS4's OS reserve for RAM).

*Specifically, 'graphics power' in the sense of visual output on screen and tech graphics performance. In other words, it was being proclaimed all over the place that PS4's 18 CU's would lead to '50% better performance' than the 12 CU's in X1.

In wholly CU limited situations that's exactly what it will do. Why do you think this not to be the case?

but the PR and common rhetoric surrounding the issue suddenly raises the specter of an apparent downgrade once ppl discover that not all of those 1.84Tflops will be used for actual graphics.

Except that in many cases they will so in fact your entire point is completely wrong.

 

[Deleted member 7537]

Driveclub creators comments on this

Kirkland is intrigued but reckons there's plenty of time to find that stuff. "You already have a lot of CPU power at your disposal," he says. "I think with many of the first generation PS4 titles, developers probably won't need to worry about it – they'll be able to get a lot out of parallelism across the CPU cores – but for teams who are a bit more ambitious, who want to do interesting things, it's just waiting there. We're doing some of that in DriveClub and I'm sure other guys will go further – and the platform guys will expose more of that functionality through the lifespan of the machine, unlocking more potential."

http://www.theguardian.com/technology/gamesblog/2013/aug/07/ps4-driveclub-future-game-visuals

Really nothing new.

 

[Bigus Dickus]

What's a "wholly CU limited" usage example? I'd love to see benchmarks that shows a 50% improvement in any actual game scenario. I know this type of benchmark was attempted, but not very well done.

 

[humbertklyka]

What's a "wholly CU limited" usage example? I'd love to see benchmarks that shows a 50% improvement in any actual game scenario. I know this type of benchmark was attempted, but not very well done.

The issue is that the relevant game scenarios will not exist until PS4 has been around for a while. Of course current games will not show such a scaling, as they would be balanced for current GPUs.

(And in reality wholly CU limited will most likely never happen, as different parts of the frame rendering will have different bottlenecks)

 

[jonabbey]

(Apologies if this is too far afield for this topic..)

While we're talking about CUs, does anyone here have a good feel for whether GPGPU on GCN can efficiently support h.265 decode?

We know the PS4 comes with hardware encode/decode for h.264, but I assume streaming services will move to h.265 not too long into the next gen consoles' lifetimes. Does the extra CUs on PS4 have any likelihood of coming into play with h.265, or is that just apples and SAT scores?

 

[Bigus Dickus]

(And in reality wholly CU limited will most likely never happen, as different parts of the frame rendering will have different bottlenecks)

That's my impression as well, but I start wondering when people talk about 50% better performance in some cases.

 

[temesgen]

Finally you said earlier in the thread that XB1 has 40% more CPU - walk me thru the math on that. I've seen that said before and I've also seen it disputed so please reply with more of detail so I can see how you arrive at that conclusion.

@Rangers - can you reply to this, genuine question... thanks

 

[Esrever]

what if the 4 CUs are actually in the cloud?

 

[zupallinere]

Oh sheeeeet!

“The most important thing is that it’s [PS4] the most powerful games console ever created,”

... or so sayeth Cerny. Kicking it up a notch on the PR front. Does this mean no GCN 2.0 CPU ??

From: http://www.gameranx.com/updates/id/16618/article/cerny-ps4-to-match-ps3-s-seven-year-cycle/

I think 7 years is gonna be a bit long. Computation will be abundant and consoles will now all be based on x86 technology for the foreseeable future. Seems to me it's about when will consumers get bored and how fast can these consoles get cost reduced and/or how much previous design work gets reused. But this is another thread.