Just keep in mind, this doesn't somehow make the PS4 any less awesome than it already is. It's how sony designed the system, and it's smart to push GPGPU.
You need to post whatever you have because you are interpreting it wrong.
Just keep in mind, this doesn't somehow make the PS4 any less awesome than it already is. It's how sony designed the system, and it's smart to push GPGPU.
It's suspiciously interesting the MS PR specifically mentioned than the GPU overclock was better than running with 14 CUs...
This isn't how the real world works, it isn't so cut and dry but I could believe it for specific scenarios.
PS4 unbalanced confirmed
that's really interesting though. And perhaps would that mean that cross-gen multiplatform titles might not show much difference between xbox and PS4? Assuming they are relatively straightforward up-ports.
That's a huge misconception. He said no different than what I said, only he danced around or left out the information communicated to devs. He actually didn't lie, he just omitted or didn't fully clarify.
In fact, Cerny greatly hinted that it was true.
http://www.eurogamer.net/articles/digitalfoundry-face-to-face-with-mark-cerny
When he says a little bit more ALU than if you were thinking strictly about graphics, and when he says the hardware is intentionally not 100 percent round, he means that it doesn't scale in returns for graphics per block of ALU like you might think.
The Graphics Processing Unit (GPU) has been enhanced in a number of ways, principally to allow for easier use of the GPU for general purpose computing (GPGPU) such as physics simulation. The GPU contains a unified array of 18 compute units, which collectively generate 1.84 Teraflops of processing power that can freely be applied to graphics, simulation tasks, or some mixture of the two.
so to paraphrase
"you could use those extra CUs for graphics, but wouldn't that be a bit of a waste? Sure your game will look a little better, but imagine how amazing it'll look if you explored using those CUs for GPGPU instead. All that amazing physics, AI, fluid dynamics - it'll look fantastic"
You need to post whatever you have because you are interpreting it wrong.
so to paraphrase
"you could use those extra CUs for graphics, but you'll hit a bit of diminishing returns - wouldn't that be a waste? Sure your game will look a little better, but imagine how amazing it'll look if you explored using those CUs for GPGPU instead. All that amazing physics, AI, fluid dynamics - it'll look fantastic"
This isn't my interpretation. It's the interpretation of an experienced games programmer that learned this at an official Sony devcon. I'm just repeating their understanding, so there's very little getting lost in translation as a result. Only way the info could be misinterpreted is if they misinterpreted it, which I doubt.
More or less, absolutely.
What's seemingly being described here seems beyond just diminishing returns - it's being made out as if improvement falls entirely off a cliff after 14 CUs - like it's the magic number, while AMD is happily churning out 16, 20, 32 CU cards etc.Well if what Gofreak said is true then it does impact desktop GPUs. He stated benchmarks of comparable GPUs with 50% more CUs only received a 25% gain in performance.
Radeon HD 7790 1GB Radeon HD 7850 1GB
BioShock Infinite 60.5 81.5
Hitman Absolution 35.5 46.7
Batman: Arkham City 65 89
Tomb Raider 53.4 69.4
Sleeping Dogs 45.9 56.3
Metro 2033 23.7 32.6
Sure. That's all well and good and seems plausible. But that's not really the only claim being made here.Sony is saying it might be better to use those extra CUs for Compute worked because you will gain a greater than x% of performance then just using them on graphics.
As above, seems completely plausible - if the extra compute resources aren't really going to do much for a given piece of software in terms of graphics it makes perfect sense to utilise them for something else otherwise it's just wasted resources. Of course, as also noted above, that doesn't seem to be the sole claim being made here - which seems to be more along the lines of the PS4 is somehow different from desktop GPUs and beyond using 14 of its CUs for graphics the performance increase plummets to the realms of negligibility.True. But they're just saying: have a look to see if there's a point where you could be using CUs better for other things.
They're asking devs to check if they're under-using ALU, holding the fairly attractive suggestion that quite a lot of power could be spare without hugely affecting render performance. But, again, it's very much your mileage may vary.
It wouldn't be pointless... like those PC benches show, you can still get a gain (variable depending on the game). In some the gain may even be linear. But if on average the gain isn't linear, and your software is being held back by some other point in the pipeline beyond a certain level of shading performance, it is simply wise to consider throwing in other work that may have a different ratio of resource demands, to get alu utilisation back up, and get some nice features or work done into the bargain.
It's not a specific impact on PS4. This is true of any GPU.
Also very true. If a multiplatform game really is utilising 400 GFLOPS of the GPU for non-graphical tasks, it probably doesn't end well for the XB1.And for sure it's in Sony's interests to push GPGPU. It could be a strong competitive advantage for them, it would be hard if not impossible to bring parity on another console if a game is chucking a decent amount of GPGPU around.
Because the way your saying it, is why even have a CPU IF part of a giant GPU can just act as one? It doesn't work that way, the GPGPU can only do certain things a CPU can and if the things needed are one of the things it can't do well or at all, then that power might go to waste.
They aren't "troublesome". It's just that you get performance penalties when using them on graphics after a certain point.
This isn't my interpretation. It's the interpretation of an experienced games programmer that learned this at an official Sony devcon. I'm just repeating their understanding, so there's very little getting lost in translation as a result. Only way the info could be misinterpreted is if they misinterpreted it, which I doubt.
so to paraphrase
"you could use those extra CUs for graphics, but you'll hit a bit of diminishing returns - wouldn't that be a waste? Sure your game will look a little better, but imagine how amazing it'll look if you explored using those CUs for GPGPU instead. All that amazing physics, AI, fluid dynamics - it'll look fantastic"
He was just making clear that performance doesn't scale linearly, which is true. He was just saying that fact might encourage devs to use that performance for compute instead of rendering. You aren't forced into anything. I don't think devs would be reporting 40-50% gains over XBO performance if 14 was the max they could use efficiently.
Cool, you know where to find me then.Only saying it now so I don't have to repeat it. I'm not posting a link, because it isn't a link. it's just information that I think proves what I'm saying is true beyond doubt, but it also carries the risk of getting someone in trouble or violating their trust, which is why I can't just say everything, and why immediately after I said it, I said I would have no issue sharing the information with a mod if it made people more comfortable that I'm being transparent and honest.
Cool, you know where to find me then.
He was just making clear that performance doesn't scale linearly, which is true. He was saying that fact might encourage devs to use that performance for compute instead of rendering. You aren't forced into anything. I don't think devs would be reporting 40-50% gains over XBO performance if 14 was the max they could use efficiently.
Also this:
http://www.scei.co.jp/corporate/release/130221a_e.html
MS must genuinely be freaking out about the negative PR coming out of their '20% but 50% less' game console considering how much they're trying to get out ahead of this, but the numbers don't lie, and the gap is pretty substantial.
Also, take a drink every time the guy refers to 'balance' in the architecture. I can only assume a small team of people at MS spent at least a few hours brainstorming a term they could use to make the situation more vague and questionable.
Just sent.
Cool, you know where to find me then.
it's being made out as if improvement falls entirely off a cliff after 14 CUs
so to paraphrase
"you could use those extra CUs for graphics, but you'll hit a bit of diminishing returns - wouldn't that be a waste? Sure your game will look a little better, but imagine how amazing it'll look if you explored using those CUs for GPGPU instead. All that amazing physics, AI, fluid dynamics - it'll look fantastic"
What exactly is being sent if what we're currently discussing is just a verbatim repetition of an interpretation of a developer presentation and not any form of official document? Like an ordinary PM :/
The whole notion of discrete distribution of CUs in this conversations is stupid though. Async compute is there so you don't have to do that.gofreak said:They're asking devs to check if they're under-using ALU, holding the fairly attractive suggestion that quite a lot of power could be spare without hugely affecting render performance.
Of course, as also noted above, that doesn't seem to be the sole claim being made here - which seems to be more along the lines of the PS4 is somehow different from desktop GPUs and beyond using 14 of its CUs for graphics the performance increase plummets to the realms of negligibility.
The whole notion of discrete distribution of CUs in this conversations is stupid though. Async compute is there so you don't have to do that.
Eg. in a deferred renderer all CUs will be effectively idle during Z-prepass(meaning ~18 would be available for compute during), and all of them could be utilized during shading pass.
Sure it may average out to 4:14 over the course of frame - but if you discretely partition it, chances are you're losing performance on both sides.
What exactly is being sent if what we're currently discussing is just a verbatim repetition of an interpretation of a developer presentation and not any form of official document? Like an ordinary PM :/
Oh well, iirc, bish is a developer as well...? So it would be interesting to get his take on this all anyway.
so to paraphrase
"you could use those extra CUs for graphics, but you'll hit a bit of diminishing returns - wouldn't that be a waste? Sure your game will look a little better, but imagine how amazing it'll look if you explored using those CUs for GPGPU instead. All that amazing physics, AI, fluid dynamics - it'll look fantastic"
That's how I understand it, yes.
And IIRC, these 4 extra CU are not the same as the other 14, they're specialized for compute which would make it an even bigger waste.
And IIRC, these 4 extra CU are not the same as the other 14, they're specialized for compute which would make it an even bigger waste.
And IIRC, these 4 extra CU are not the same as the other 14, they're specialized for compute which would make it an even bigger waste.
All 18 CUs are the same.
The whole notion of discrete distribution of CUs in this conversations is stupid though. Async compute is there so you don't have to do that.
Eg. in a deferred renderer all CUs will be effectively idle during Z-prepass(meaning ~18 would be available for compute during), and all of them could be utilized during shading pass.
Sure it may average out to 4:14 over the course of frame - but if you discretely partition it, chances are you're losing performance on both sides.
Keep in mind I didn't say it couldn't be used however a dev wished. Just what Sony presented to devs about a bit of diminishing returns for graphics operations after the 14 CU mark. Devs if they wish can use all 18 for graphics, but apparently there's a significant drop off in the benefit that you see for the additional ALU resources for graphics past the 14 CU mark, which is why Sony suggests the most optimal use for the extra 4 CUs is to use them for compute work.
"The exact information was that given the rest of the design there is a huge knee in the performance curve, and anything beyond that point there is a significant drop off in the apparent value that you get from additional ALU resources for graphics, and the PS4 is said to be well to the right of that knee.
And IIRC, these 4 extra CU are not the same as the other 14, they're specialized for compute which would make it an even bigger waste.
I'm glad you learned your lesson SenjutsuSage. Spinning makes your head dizzy.
But that doesn't make sense, that 14 is some kind of magic number. It will completely depend on your game engine. That's like someone saying "after 6 GB of RAM you have diminishing returns in a PC". Well, sure, maybe if you're only using it to read your email, but if you're running a DB on that machine you'l easily soak up the 6GB.
And you didn't even claim diminishing returns, originally, you said this:
Mr Sage isn't banned yet. Docs real?
Mr Sage isn't banned yet. Docs real?
Nothing about diminishing returns was ever stated by anyone but ms pr guys. the changes made by sony would not be needed if you were always have some cu just for gpgpu work. Not buying any of this nonsense.Just sent.
Keep in mind I didn't say it couldn't be used however a dev wished. Just what Sony presented to devs about a bit of diminishing returns for graphics operations after the 14 CU mark. Devs if they wish can use all 18 for graphics, but apparently there's a significant drop off in the benefit that you see for the additional ALU resources for graphics past the 14 CU mark, which is why Sony suggests the most optimal use for the extra 4 CUs is to use them for compute work.
Yes but distinction is important when arguing about "wasted CU potential in graphics", as scenarios where you target max consumption at different parts of frame are quite viable in closed boxes, even if bandwith/fill suggests you can't in average case.gofreak said:Sure, but it's just easier to talk about in terms of discrete CUs to try and quantify the amount of performance that could be going a-begging.
No. Games are inherently spiky when it comes to performance, nothing is better off with evenly distributed average, that's the entire reason why we have unified shaders since 2005.NTIETS13A_Superiority said:And wouldnt constant, say 4 CUs for compute better as you have more reliable access to resources?
I did
That explains exactly diminishing returns. There is a drop off in the value of the extra ALU resources for graphics after a certain point, that point being 14 CUs.
That explains exactly diminishing returns. There is a drop off in the value of the extra ALU resources for graphics after a certain point, that point being 14 CUs.
And that is complete BS. It depends on application/tasks. I forgot we have ignorelist.
And IIRC, these 4 extra CU are not the same as the other 14, they're specialized for compute which would make it an even bigger waste.
Hopefully bish gives us his interpretation of it as I think there is some truth to what Senjutsu said but likely not quite as big a deal as originally suggested
Curious to know what the returns are like from those 4 CU's
Although not sure I'd understand it to be honest
And that is complete BS. It depends on application/tasks. I forgot we have ignorelist.