Support NeoGAF

[kyliethicc]

Well I'm not assuming that; hence why I said "whether they are currently being utilized."

That's really the room XSX would have to grow it's performance over the PS5; if they currently are being utilized and the system is still only on-Par with PS5 then it doesn't have that room to grow.

PS5 of course likely has room to grow performance w/ it's more custom features; but we also don't know for sure how much of that is currently being used or not, or for which games. Same with XSX; we just don't have the details.

Sure but more shaders is not going to automatically result in any better performance, even if fully utilized... if, for example, the system has split memory bandwidth bottlenecking it.

Or if the system with less CUs has a GPU clocked much faster and has less CPU cache latency. Or if that system can feed its RAM faster due to a faster I/O, etc.

Just because a system has more TFLOPs doesn't mean it'll be running games at higher res/fps, even if dem flops are fully used.

 

[MasterCornholio]

I don't believe or disbelieve personally.

Nobody but developers who have done profiling of their own games across both systems knows.

It's not like they don't have access to those additional CUs.

 

[SlimySnake]

I like Cerny, but you guys hang on every word like he isn't someone marketing his product first and foremost.

I dont know. I thought Cerny was bullshitting us back then but the final product and the numbers proved him right.

Here is a post of mine from this thread back when the specs were first revealed in March. And I wasnt the only one either. Pretty much everyone who replied to me agreed in one form or another that the PS5 wasnt well designed.

Just thinking about this. Cerny fucked this up so much that he ended up proving VFXVeteran and Timdog right. He designed the console like these github truthers wanted him too. Narrow and fast and inefficient.

If that's not the most ironic display of incompetence i dont know what is.

I have more embarrassing posts making me look like an idiot if you want to see more receipts.

 

[ethomaz]

different architectures. vega 56 doesnt outperform vega 64 or vega 7 for example.

Exactly.
PS5 and Series X are way different.
One is machine focused in efficiency while the other is a machine focused high TF count.

PS5 is similar of what nVidia did with Maxwell but for console space.

 

[Rudius]

XSS version is 60 fps, the XSX graphics mode is 30 fps.

Thats how bad any 30 fps mode looks in comparison to 60 fps.

I tried to play at 30fps on PS5, but with the possibility of 60 I just couldn't.
I don't understand why didn't they offered a lower resolution with ray-tracing at 60; 1080p would probably be enough, even if it had less stable framerates.

 

[IntentionalPun]

Just because a system has more TFLOPs doesn't mean it'll be running games at higher res/fps, even if dem flops are fully used.

So if the same game is using ~40CU's today on the XSX and it's patched to use ~50CU's, you saying it wouldn't necessarily perform better? (total rhetorical scenario of course)

 

[kyliethicc]

I dont know. I thought Cerny was bullshitting us back then but the final product and the numbers proved him right.

Here is a post of mine from this thread back when the specs were first revealed in March. And I wasnt the only one either. Pretty much everyone who replied to me agreed in one form or another that the PS5 wasnt well designed.



I have more embarrassing posts making me look like an idiot if you want to see more receipts.

 

[IntentionalPun]

I dont know. I thought Cerny was bullshitting us back then but the final product and the numbers proved him right.

I'm not saying he IS bullshitting; I'm saying that no matter what he's going to highlight the differences between his machine and the competition and say the way he's doing it is better. He did this w/o mentioning them, but the references were rather obvious.

He may be right of course. We won't REALLY know until a couple years into the gen.

The PS5 is fantastic either way.. and I'm sure people playing games on XSX aren't actually noticing many framerate dips. Games are still performing generally near 60.

 

[kyliethicc]

So if the same game is using ~40CU's today on the XSX and it's patched to use ~50CU's, you saying it wouldn't necessarily perform better? (total rhetorical scenario of course)

If all things in systems A and B are the same, except system B has more shaders, (and they can be used by the game's code,) I'd assume it'd result in an increase in gaming performance. How much? Idk. There seems to be clear diminishing returns to increasing shader counts in GPUs.

Thing is with PS5 and XSX... all other things are not the same. The system's have many differences besides just CUs.

 

[IntentionalPun]

If all things in systems A and B are the same, except system B has more shaders, (and they can be used by the game's code,) I'd assume it'd result in an increase in gaming performance. How much? Idk. There seems to be clear diminishing returns to increasing shader counts in GPUs.

Thing is with PS5 and XSX... all other things are not the same. The system's have many differences besides just CUs.

I'm not comparing PS5 to XSX there.

I'm talking about the same game, on XSX, that via a patch, increases CU utilization. (or let's say.. the sequel to that game.. or a new version of the engine used by that game, etc.,etc.)

 

[SlimySnake]

I think Cerny HAD to make those comparisons because we were all under the impressions that all tflops are the same. When in fact, higher clocks almost always resulted in better performance. We saw this with Nvidia 10 series and 20 series card hitting really high clocks compared to AMD's polaris and vega offerings and it wasnt till the 5000 series card that AMD finally started offering equal performance per flop. 5700xt and 2070 were the first cards to offer roughly the same performance per flop because AMD was finally able to hit the higher clocks. We now see AMD outperform the 30 series cards in standard rasterization performance with far fewer CUs or shader processors. The 3080 for instance has 138 CUs worth of shader processors and the 72 CU 6800xt is able to keep up with it in most non RT games thanks to the higher clocks.

RDNA 2 was a paradigm shift and tbh, no one here thought that AMD would be able to offer better performance with far fewer CUs. Yes the infinity cache helps but we are seeing the PS5 keep up the XSX its clear that higher clocks are far more important than more CUs or shader cores/processors. And since AMD hadnt revealed RDNA 2 back then, Cerny had to be the one to break this to everyone.

 

[kyliethicc]

I'm not comparing PS5 to XSX there.

I'm talking about the same game, on XSX, that via a patch, increases CU utilization. (or let's say.. the sequel to that game.. or a new version of the engine used by that game, etc.,etc.)

Sure? But why would a released game not be using all 3328 shaders on XSX right now? Unless its old back compat 360 games or some shit.

But like Control XSX version? It's definitely using all 52 CUs. I can't imagine why it would only be using like 40 CUs. Plenty of PC GPUs run Control using far more shaders.

 

[kyliethicc]

.. we were all under the impressions that all tflops are the same.

nah .. just some of us lol

 

[IntentionalPun]

Sure? But why would a released game not be using all 3328 shaders on XSX right now? Unless its old back compat 360 games or some shit.

I don't know; why did Cerny claim CU occupancy was an issue for wider GPUs? Unless I'm misremembering.

 

[kyliethicc]

I don't know; why did Cerny claim CU occupancy was an issue for wider GPUs? Unless I'm misremembering.

No he said that, and since I don't code, idk fully. I got the impression he was just saying there are some diminishing returns to using more shaders. That its easier to get the most out of 36 CUs than it is to get the most out of 48 CUs. (But in that example, all 48 CUs are still technically being used, just perhaps inefficiently.)

He didn't imply its impossible, just more difficult. Maybe depends on the game, the engine, the programmers, APIs, dev timeframe, etc?

I think this is why Cerny kept PS5 with same 8 CPU cores and 36 CUs as PS4 Pro. Seems he's hoping it'll be easier for devs to use it all as fully as possible. (Among other reasons for the system design.)

 

[MasterCornholio]

I don't know; why did Cerny claim CU occupancy was an issue for wider GPUs? Unless I'm misremembering.

But XSX isn't exactly wider with the same number of shader arrays as the PS5.

 

[IntentionalPun]

No he said that, and since I don't code, idk fully. I got the impression he was just saying there are some diminishing returns to using more shaders. That its easier to get the most out of 36 CUs than it is to get the most out of 48 CUs. (But in that example, all 48 CUs are still technically being used, just perhaps inefficiently.)

He didn't imply its impossible, just more difficult. Maybe depends on the game, the engine, the programmers, APIs, dev timeframe, etc?

I think this is why Cerny kept PS5 with same 8 CPU cores and 36 CUs as PS4 Pro. Seems he's hoping it'll be easier for devs to use it all as fully as possible. (Among other reasons for the system design.)

Gotcha; that comment of his is the basis for all of my comments here.. that's what I've been getting at.

I'm merely suggesting that based on what Cerny said, perhaps over time CU occupancy would improve on XSX.. but maybe it already isn't an issue.. I don't know.

 

[Alex Scott]

I don't know; why did Cerny claim CU occupancy was an issue for wider GPUs? Unless I'm misremembering.

There are things things that teraflops don't tell you. How fast the the GPU looks up texture, how are triangles are assigned to CUs. These have more impact on performance than higher teraflop. Also the most expensive thing GPU does is accessing the data not rendering i.e doing the math.

TERAFLOPS IS A GOOD METRIC FOR TELLING YOU NON-GRAPHIC WORKLOADS WHERE IT MAKES SENSE.

Backward compatible games will not push these consoles to it's edge. It is best to wait for next gen games only to see how these consoles compare.

 

[IntentionalPun]

There are things things that teraflops don't tell you. How fast the the GPU looks up texture, how are triangles are assigned to CUs. These have more impact on performance than higher teraflop. Also the most expensive thing GPU does is accessing the data not rendering i.e doing the math.

TERAFLOPS IS A GOOD METRIC FOR TELLING YOU NON-GRAPHIC WORKLOADS WHERE IT MAKES SENSE.

Backward compatible games will not push these consoles to it's edge. It is best to wait for next gen games only to see how these consoles compare.

Why are you shouting this at me?

I'm talking specifically about whether the XSX will improve it's use of teraflops. Comparing XSX of today to XSX of tomorrow, if CUs aren't being utilized now and are later.

We've been over this for a year lol, and I always thought what Cerny said had merit, and wasn't shocked when the PS5 performed as well/better than XSX.

I'm talking about XSX's potential room to grow performance; partly based on Cerny talking about CU utilization.

Nobody in this convo needs a "babbys first Teraflop convo."

 

[TheThreadsThatBindUs]

But again, the difference between 70 series Nvidia cards and 80 series cards is usually 18-20% and that always translates into 18-20% better performance. Same goes for the 5700 and 5700xt and 6800 and 6800xt.

Well in this case you're comparing two GPUs.

In the console's case you're comparing two entire systems with different micro and macro architectural design choices and customizations

The question about XSX is whether the extra CUs are currently being utilized, or if they aren't, if they ever will be.

Well if one box is designed to improve cache hit rate then by definition the PS5's CUs will be better utilized.

99% of GPU performance is ensuring those execution units are sufficiently fed with data, and to do that you need to make sure the data needed is present in the caches when it's needed.

I like Cerny, but you guys hang on every word like he isn't someone marketing his product first and foremost.

Or... people are seeing the proof in the pudding that the PS5 is consistently outperforming the XSX and they're concluding that Cerny's PS5 customizations and design choices are benefitting the system more than what MS did.

 

[IntentionalPun]

Or... people are seeing the proof in the pudding that the PS5 is consistently outperforming the XSX and they're concluding that Cerny's PS5 customizations and design choices are benefitting the system more than what MS did.

Right the PS5 is performing fantastically. I'm just talking about whether the XSX has room to grow or not.. and basing what I'm saying, on Cerny's own comments about CU utilization.

If the CU utilization will never get better on XSX due to less optimized data pipelines then so be it.

But surely if it's CU utilization improves (and isn't great now, maybe it is), it will perform better than it is today.

That's all I'm suggesting.

None of this was comparing the XSX to PS5.... it was comparing the XSX of today to a potential future more utilized XSX.

Now let me go and play with 1 of my 2 PS5s lol

 

[SlimySnake]

But XSX isn't exactly wider with the same number of shader arrays as the PS5.

Do we have block diagrams for the new 6800xt cards. Did they do the same thing as MS and stick in extra CUs in the two existing shader engines or did they add more?

IIRC, the limit used to be 20 CUs per shader engine but with RDNA the concept of shader engines went away and even though AMD didnt release any cards with more than 20 per shader array, MS bolted on a bunch of extra CUs to get an advantage on paper. Maybe thats the bottleneck here?

XSX has enough vram bandwidth and at least enough vram for cross gen games like Hitman, CoD, AC and Control. The 3080 has 10 GB of vram and it seems to do fine even at 4k 120 fps for games like Doom. So if there is a bottleneck in the split ram structure, it shouldnt affect cross gen games. their CPU is pretty much identical so the bottleneck must be the GPU.

 

[Alex Scott]

Why are you shouting this at me?

I'm talking specifically about whether the XSX will improve it's use of teraflops. Comparing XSX of today to XSX of tomorrow, if CUs aren't being utilized now and are later.

We've been over this for a year lol, and I always thought what Cerny said had merit, and wasn't shocked when the PS5 performed as well/better than XSX.

I'm talking about XSX's potential room to grow performance; partly based on Cerny talking about CU utilization.

Nobody in this convo needs a "babbys first Teraflop convo."

Shouting wasn't my intention. I should have highlighted my main point instead.

I am sure developers will continue to push the consoles to get the best out of it.

But if you are applying that current games will improve their performance 2 years from now because of better CUs utilization then I don't think so. Neither will PS5.

 

[LiquidRex]

Keep an eye out for RGT video on VRS, will be very interesting... Subscribe if you haven't, this channel is going places. ☺

 

[roops67]

Now let me go and play with 1 of my 2 PS5s lol

Just rub it in why don't you! You know there's many of us here that still can't get our hands on one

 

[IntentionalPun]

Just rub it in why don't you! You know there's many of us here that still can't get our hands on one

Well FWIW I'm playing that Fenyx Rising game and it looks / runs like a PS4 game lol

 

[FrankWza]

[

 

[bitbydeath]

Why is it that every PlayStation console generation has had multiple Tech demo's apart from the PS5 which only had one. (UE5) don't think we can blame Covid, is it just a case UE5 was more than enough to showcase the consoles potential.

We did get this one as well, although it wasn’t labeled at the time as a next-gen demo.



That’s the new engine that’s running Project Athia.

 

[kyliethicc]

Gotcha; that comment of his is the basis for all of my comments here.. that's what I've been getting at.

I'm merely suggesting that based on what Cerny said, perhaps over time CU occupancy would improve on XSX.. but maybe it already isn't an issue.. I don't know.

For some games I'm sure it is, some probably not. Will that change over time? sure. But probably not for a released game.

Like Assassin's Creed Vahalla. That game is never gonna suddenly run better. Its done and shipped.

But could Ubisoft devs get better use out of lots of shaders in future AC games? I'd assume yes. Will they? Idk.

 

[IntentionalPun]

For some games I'm sure it is, some probably not. Will that change over time? sure. But probably not for a released game.

Like Assassin's Creed Vahalla. That game is never gonna suddenly run better. Its done and shipped.

But could Ubisoft devs get better use out of lots of shaders in future AC games? I'd assume yes. Will they? Idk.

I was just using that as a rhetorical scenario, not a realistic one, to make it clear what I was getting at.

And as I said.. PS5 has features potentially being underutilized.. as I've been telling some of the Xbots here.. don't be shocked if it's perf improves over time more than XSX.

I own both systems (although might never use my XSX TBH) so I hope both improve.

 

[kyliethicc]

Do we have block diagrams for the new 6800xt cards. Did they do the same thing as MS and stick in extra CUs in the two existing shader engines or did they add more?

IIRC, the limit used to be 20 CUs per shader engine but with RDNA the concept of shader engines went away and even though AMD didnt release any cards with more than 20 per shader array, MS bolted on a bunch of extra CUs to get an advantage on paper. Maybe thats the bottleneck here?

XSX has enough vram bandwidth and at least enough vram for cross gen games like Hitman, CoD, AC and Control. The 3080 has 10 GB of vram and it seems to do fine even at 4k 120 fps for games like Doom. So if there is a bottleneck in the split ram structure, it shouldnt affect cross gen games. their CPU is pretty much identical so the bottleneck must be the GPU.

Navi 10 die aka 5700 XT is 5 Dual Compute Units per Shader Array. 2 arrays per shader engine. 2 shader engines.

Navi 21 die aka 6900 XT / 6800 XT is same 5 DCUs per array. But 4 shader engines total. Thats why they called it Navi 2X.

PS4 & PS5 follow the same core layout as Navi GPUs, but Xbox do not.

RDNA1 NAVI 10 - 5 DCUs per array (5x2x2x2)
RDNA2 NAVI 21 - 5 DCUs per array (5x2x2x4)

PS4 - 5 DCUs per array (5x2x2x1)
PS4K - 5 DCUs per array (5x2x2x2)
PS5 - 5 DCUs per array (5x2x2x2)

XOne - 7 DCUs per array (7x2x1x1)
XOneX - 11 DCUs per array (11x2x2x1)
XSX - 7 DCUs per array (7x2x2x2)
XSS - 6 DCUs per array (6x2x2x1)

 

[TheThreadsThatBindUs]

Right the PS5 is performing fantastically. I'm just talking about whether the XSX has room to grow or not.. and basing what I'm saying, on Cerny's own comments about CU utilization.

If the CU utilization will never get better on XSX due to less optimized data pipelines then so be it.

But surely if it's CU utilization improves (and isn't great now, maybe it is), it will perform better than it is today.

That's all I'm suggesting.

None of this was comparing the XSX to PS5.... it was comparing the XSX of today to a potential future more utilized XSX.

Now let me go and play with 1 of my 2 PS5s lol

Cerny's comments about CU utilization are a little cryptic. We know GPU graphics workloads generally tend to be embarrassingly parallelizable. That said, you will struggle to keep CUs occupied if you can't feed them with data fast enough. And so I think that this is what Cerny was intimating with his explanations around a "fast narrow vs wide slower" design.

With a narrow faster GPU, your front end is faster, the cache bandwidth is higher, and so it's easier to keep your fewer CUs fed and thus occupied.

With a wide slower design, your front end is slower, cache bandwidth is lower, and so even though you have a wider shading array, it's harder to keep the CUs fed because the infrastructure that feeds the CUs is performing disproportionately worse.

So I don't think it's a developer issue, or API/driver issue that can be improved over time. It's something that is essentially embedded into the hardware design itself.

 

[MasterCornholio]

Cerny's comments about CU utilization are a little cryptic. We know GPU graphics workloads generally tend to be embarrassingly parallelizable. That said, you will struggle to keep CUs occupied if you can't feed them with data fast enough. And so I think that this is what Cerny was intimating with his explanations around a "fast narrow vs wide slower" design.

With a narrow faster GPU, your front end is faster, the cache bandwidth is higher, and so it's easier to keep your fewer CUs fed and thus occupied.

With a wide slower design, your front end is slower, cache bandwidth is lower, and so even though you have a wider shading array, it's harder to keep the CUs fed because the infrastructure that feeds the CUs is performing disproportionately worse.

So I don't think it's a developer issue, or API/driver issue that can be improved over time. It's something that is essentially embedded into the hardware design itself.

Very interesting. But what if instead of going wider you just add more CUs to the shader arrays?

 

[kyliethicc]

Cerny's comments about CU utilization are a little cryptic. We know GPU graphics workloads generally tend to be embarrassingly parallelizable. That said, you will struggle to keep CUs occupied if you can't feed them with data fast enough. And so I think that this is what Cerny was intimating with his explanations around a "fast narrow vs wide slower" design.

With a narrow faster GPU, your front end is faster, the cache bandwidth is higher, and so it's easier to keep your fewer CUs fed and thus occupied.

With a wide slower design, your front end is slower, cache bandwidth is lower, and so even though you have a wider shading array, it's harder to keep the CUs fed because the infrastructure that feeds the CUs is performing disproportionately worse.

So I don't think it's a developer issue, or API/driver issue that can be improved over time. It's something that is essentially embedded into the hardware design itself.

He did add the qualifier "when triangles are small" or something like that.

Its probably related to shit he and or other Sony programmers have found when making PS4 1st party games.

I bet a fair amount of how the PS5 got designed was from SIE WWS dev feedback on what would help them make their exclusive games.

 

[TheThreadsThatBindUs]

Very interesting. But what if instead of going wider you just add more CUs to the shader arrays?

Adding more CUs is what we mean when we say "make the design wider".

He did add the qualifier "when triangles are small" or something like that.

Its probably related to shit he and or other Sony programmers have found when making PS4 1st party games.

I bet a fair amount of how the PS5 got designed was from SIE WWS dev feedback on what would help them make their exclusive games.

Oh, that makes a lot of sense too.

We know smaller triangles reduce rendering efficiency, because it quickly reduces ALU utilization. So when rendering a scene with many smaller triangles, adding more ALUs (i.e. wider shading array / more CUs) won't make as much of a difference to rendering performance as increasing clock speed.

In Cerny's own words, "a rising tide lifts all boats".

 

[IntentionalPun]

Smaller triangles sounds like UE5.

 

[TheThreadsThatBindUs]

Smaller triangles sounds like UE5.

Yeah Nanite is cool because it gets around this rigid hardware limitation of smaller triangles reducing ALU utilization, by doing all the primitive assembly, culling, parameterization and most importantly rasterization in software.

So with Nanite, the result of the whole "fast and Narrow versus slow and wide" design dichotomy with scenes with smaller tris will look very different than the traditional graphics rendering path.

 

[kyliethicc]

Smaller triangles sounds like UE5.

Yup. And other various AAA games with fancy graphics and lots of polygons. Basically most of Sony's AAA games over the last gen.

Like games with detailed realistic character models made up of fuck tons of polygons (Last of Us 2, God of War, Death Stranding, Uncharted, etc)

 

[Rea]

Adding more CUs is what we mean when we say "make the design wider".



Oh, that makes a lot of sense too.

We know smaller triangles reduce rendering efficiency, because it quickly reduces ALU utilization. So when rendering a scene with many smaller triangles, adding more ALUs (i.e. wider shading array / more CUs) won't make as much of a difference to rendering performance as increasing clock speed.

In Cerny's own words, "a rising tide lifts all boats".

I don't understand why" when triangles are small its harder to fill CUs with meaningful work." Quote. How does The size of the triangles has directly related to CU counts?

 

[Sinthor]

different architectures. vega 56 doesnt outperform vega 64 or vega 7 for example.

But again, the difference between 70 series Nvidia cards and 80 series cards is usually 18-20% and that always translates into 18-20% better performance. Same goes for the 5700 and 5700xt and 6800 and 6800xt.

I am not saying I disagree with you because I do believe the PS5's higher clocks, unified ram pool and custom APU is mitigating the tflops gap. Hitman of course not withstanding.

I think it's overall bandwidth (not memory bandwidth, but I mean the system overall) and the caching of the PS5 that is causing these results. I also believe the XSX split memory configuration is definitely impacting the system in performance. In fact, I think it's likely that that is the majority of the reason for these performance results we're seeing. I think we'll be able to tell more as we have more native games to play. If the XSX does have an advantage, those games should show it. If we still are seeing the same results (possibly even worse as devs start using the Geometry Engine and other features supposedly) then I think we'll know. The only major issue I see right now is that these results are BC games and not native. I think we need to see what a slew of truly next gen native games show us before we draw final conclusions.

 

[TheThreadsThatBindUs]

I don't understand why" when triangles are small its harder to fill CUs with meaningful work." Quote. How does The size of the triangles has directly related to CU counts?

Because with rasterization, triangles/primitives are mapped to the pixel grid before pixel shading occurs. The ALUs (Arithmetic and Logic Units) within the CUs work on these portions of the pixel grid for each polygon by breaking it further down into pixel fragments and passing the work to ALU to work on it on a pixel-by-pixel basis.

To ensure the maximum level of parallelism, in order to keep you ALU doing useful work, you want triangles that span a large number of pixels. So that when shading, in batches of pixels sized to correspond to the width of the ALUs, you're maximizing utilization of your ALU.

A crude example is as follows:

Let's say I have a 4x4 fragment passed to my 16-wide SIMD unit within the CU.

If the polygon spans 14 of the 16 pixels, then 14 out of 16 pixels get shaded in a single clock cycle (87.5% utilization)
If the polygon is smaller and spans only 4 of the 16 pixels, only 4 pixels get shaded per clock cycle (25% utilization) so the efficiency is significantly lower.

So in the above example scene with many 4x pixel triangles, adding more CUs to my GPU barely helps real world shading performance at all, because I'm getting only 25% utilization out of my ALUs. Increasing GPU clock speed however, will help overall performance more, because in a 30fps game with 33ms frame time, increasing my clock speed will mean I can shade more real 4x pixel polygons within my frame-time budget.

The above is kind of a gross oversimplification, but it's just to give you the gist.

 

[MistBreeze]

I think it's overall bandwidth (not memory bandwidth, but I mean the system overall) and the caching of the PS5 that is causing these results. I also believe the XSX split memory configuration is definitely impacting the system in performance. In fact, I think it's likely that that is the majority of the reason for these performance results we're seeing. I think we'll be able to tell more as we have more native games to play. If the XSX does have an advantage, those games should show it. If we still are seeing the same results (possibly even worse as devs start using the Geometry Engine and other features supposedly) then I think we'll know. The only major issue I see right now is that these results are BC games and not native. I think we need to see what a slew of truly next gen native games show us before we draw final conclusions.

exactly ... Im with u

split memory pool is always a bad Idea remember ps3 ?? it was one of the main advantages for xbox 360 to have unified memory

but matter of fact that sony successfully made a machine gives same performance as series x with less Teraflops

is quite an achievement

 

[SlimySnake]

exactly ... Im with u

split memory pool is always a bad Idea remember ps3 ?? it was one of the main advantages for xbox 360 to have unified memory

but matter of fact that sony successfully made a machine gives same performance as series x with less Teraflops

is quite an achievement

I think it's overall bandwidth (not memory bandwidth, but I mean the system overall) and the caching of the PS5 that is causing these results. I also believe the XSX split memory configuration is definitely impacting the system in performance. In fact, I think it's likely that that is the majority of the reason for these performance results we're seeing. I think we'll be able to tell more as we have more native games to play. If the XSX does have an advantage, those games should show it. If we still are seeing the same results (possibly even worse as devs start using the Geometry Engine and other features supposedly) then I think we'll know. The only major issue I see right now is that these results are BC games and not native. I think we need to see what a slew of truly next gen native games show us before we draw final conclusions.

But the 3080 has only 10 gb of vram and has no issues destroying the xsx in performance in these same games. control is a last gen game and is likely using only 5GB of vram at any given moment. Hitman was definitely using only 4.5 gb of vram on PCs at native 4k.

The PS3 only had 256MB of VRAM. it wasnt just because it was split, they just didnt have enough vram. xbox series x's vram should be enough unless they have somehow managed to fuck it up and the smaller slower pool of ram is slowing down the faster pool of ram.

 

[SlimySnake]

Smaller triangles sounds like UE5.

Sure, but are they being culled?

 

[TheThreadsThatBindUs]

I think it's overall bandwidth (not memory bandwidth, but I mean the system overall) and the caching of the PS5 that is causing these results. I also believe the XSX split memory configuration is definitely impacting the system in performance. In fact, I think it's likely that that is the majority of the reason for these performance results we're seeing. I think we'll be able to tell more as we have more native games to play. If the XSX does have an advantage, those games should show it. If we still are seeing the same results (possibly even worse as devs start using the Geometry Engine and other features supposedly) then I think we'll know. The only major issue I see right now is that these results are BC games and not native. I think we need to see what a slew of truly next gen native games show us before we draw final conclusions.

I'm not sure the bolded is a safe assumption or a safe conclusion to draw.

Given we're currently in the era of cross-gen where most games are designed with memory budgets to accommodate last-gen's measly 5-6GB available RAM for games. XSX's fast pool is 10GB and I believe most if not all of this is available to the game.

When we move to more native games made exclusively for next-gen, where devs will be pushing well up against the limits of available RAM on the new-gen machines, I suspect the situation could be made worse versus PS5, in cases where devs want to use more than the 10GB available in the XSX fast pool; where they'd have to burn many GPU clock cycles switching over to access the slower pool on the fly.

On PS5 they will have slightly lower overall bandwidth than XSX (but not relative to the GPU perf.), but a bigger pool available than 10GB that they will be able to uniformly access at the full speed.

Either way, we'll see.

 

[kyliethicc]

But the 3080 has only 10 gb of vram and has no issues destroying the xsx in performance in these same games. control is a last gen game and is likely using only 5GB of vram at any given moment. Hitman was definitely using only 4.5 gb of vram on PCs at native 4k.

The PS3 only had 256MB of VRAM. it wasnt just because it was split, they just didnt have enough vram. xbox series x's vram should be enough unless they have somehow managed to fuck it up and the smaller slower pool of ram is slowing down the faster pool of ram.

A guy in the control thread posted a video of him playing on a 2080 max Q laptop. Game was 1440p with RT (I think) and using around 9 GB VRAM and 9 GB RAM, so 18 GB RAM total.

Consoles use all of their RAM. Its always a limiting factor. And Xbox splitting their RAM & bandwidth hurts performance. The lead programmer of Doom Eternal said its a major issue.

 

[Shmunter]

But the 3080 has only 10 gb of vram and has no issues destroying the xsx in performance in these same games. control is a last gen game and is likely using only 5GB of vram at any given moment. Hitman was definitely using only 4.5 gb of vram on PCs at native 4k.

The PS3 only had 256MB of VRAM. it wasnt just because it was split, they just didnt have enough vram. xbox series x's vram should be enough unless they have somehow managed to fuck it up and the smaller slower pool of ram is slowing down the faster pool of ram.

The difference may be in that the 3080 has 10gig at it's disposal. On XsX it may be that when the slow pool of data is continually accessed by any component...e.g. the cpu, the entire bandwidth gets downgraded during the transfer and affecting gpu bandwidth.

There may be ways to optimize around it with some data striping and clever access, but it may be a headache many don't bother to deal with.

I mean it's all speculation, but when looking for answers - this is the difference between the systems and sticks out. What else can we point to? Could be a reason it's never been attempted before (ps3 is different)

 

[Sinthor]

I'm not sure the bolded is a safe assumption or a safe conclusion to draw.

Given we're currently in the era of cross-gen where most games are designed with memory budgets to accommodate last-gen's measly 5-6GB available RAM for games. XSX's fast pool is 10GB and I believe most if not all of this is available to the game.

When we move to more native games made exclusively for next-gen, where devs will be pushing well up against the limits of available RAM on the new-gen machines, I suspect the situation could be made worse versus PS5, in cases where devs want to use more than the 10GB available in the XSX fast pool; where they'd have to burn many GPU clock cycles switching over to access the slower pool on the fly.

On PS5 they will have slightly lower overall bandwidth than XSX (but not relative to the GPU perf.), but a bigger pool available than 10GB that they will be able to uniformly access at the full speed.

Either way, we'll see.

Definitely agree it's not a safe assumption. Just speculation on my part, for sure. Like I was saying....I won't be comfortable in even the comparison results till we see more from next gen 'native' games. These BC games are close for sure but the varied results we're seeing don't stack up really well. They seem too...all over the map. I definitely think we will see more realistically if there is a difference in performance and to what degree when we have more 1st party titles to evaluate and some truly designed for next gen 3rd party games we can compare as well.

What I mean about the memory is that there are some seemingly legit 'experts' claiming than ANY access of the slower memory on the XSX slows down the whole pipeline. Some are even saying that since OS level features like quick resume are present that this is happening regularly even if the game itself doesn't need that RAM. I'm not an expert like that, but those explanations do seem feasible in explaining the results seen. Except for, again, I think we need native built code to compare and not all this BC.

So yes, just speculation and fun speculation but it's going to be some time I think before we can confidently say one way or the other if one of these next gen consoles truly has an advantage or if performance is pretty much equal. For now, a lot of these results are just puzzling. It REALLY makes me wish I WAS an EE design expert or something like that!

 

[kyliethicc]

The difference may be in that the 3080 has 10gig at it's disposal. On XsX it may be that when the slow pool of data is continually accessed by any component...e.g. the cpu, the entire bandwidth gets downgraded during the transfer and affecting gpu bandwidth.

There may be ways to optimize around it with some data striping and clever access, but it may be a headache many don't bother to deal with.

I mean it's all speculation, but when looking for answers - this is the difference between the systems and sticks out. What else can we point to? Could be a reason it's never been attempted before (ps3 is different)

Makes perfect sense.

Only other things it could really be besides RAM is what? CPU?

 

[Sinthor]

I dont know. I thought Cerny was bullshitting us back then but the final product and the numbers proved him right.

Here is a post of mine from this thread back when the specs were first revealed in March. And I wasnt the only one either. Pretty much everyone who replied to me agreed in one form or another that the PS5 wasnt well designed.



I have more embarrassing posts making me look like an idiot if you want to see more receipts.

Hey, kudos to you, buddy! I don't recall POSTING anything like that, but you definitely weren't alone in THINKING it. I felt like Sony fucked up by the numbers for sure...trying to be too price conscious and possibly going for more exotic solutions that left them short in the performance area. I too thought the March reveal was mostly bullshit. Yeah, I know that even by the numbers the difference wasn't HUGE (18% or whatever) but it definitely seemed like the XSX would have a noticeable advantage.

So I've been pleasantly surprised to say the least, even though I still feel like we need more of those native games to really tell us where things will be.

Again, kudos on manning up to what you posted. Seriously. That's a kind of confidence, courage or whatever that's rare these days. So good for you!

Now...let's just all hope that when we GET all those native games the tables don't turn somehow to where we're wanting to throw ourselves out a window again thinking that Sony fucked up!