Current Generation Games Analysis Technical Discussion [2022] [XBSX|S, PS5, PC]

[iroboto]

Imo this memory impct is as far fatched as more popular statement that wider gpu is better with rt (ignoring clock and tf number for some reason)

Though I’m pretty sure if you run a regression on all GPUs with ray tracing, you will find the wider GPU has superior performance as a whole across the board. The only exception to this rule is actually XSX.

the trend line for wider GPU performance is correct there’s no need to graph it when each family of GPU the most performance comes with the largest chip. The cost savings methods on cut down rops, awkward memory setup combined with slower locked clocks that XSX has makes it an outlier to this trend.

if you change perspective for a bit, XSX could be viewed as a narrow chip with a wide compute array. As opposed to a wide chip as a whole. From an architectural perspective this matches more than what a wide 6000 series GPU actually looks like.

 

[snc]

Though I’m pretty sure if you run a regression on all GPUs with ray tracing, you will find the wider GPU has superior performance as a whole across the board. The only exception to this rule is actually XSX.

the trend line for wider GPU performance is correct there’s no need to graph it when each family of GPU the most performance comes with the largest chip. The cost savings methods on cut down rops, awkward memory setup combined with slower locked clocks that XSX has makes it an outlier to this trend.

if you change perspective for a bit, XSX could be viewed as a narrow chip with a wide compute array. As opposed to a wide chip as a whole. From an architectural perspective this matches more than what a wide 6000 series GPU actually looks like.

Its impossible to scale above teoretical tflops advantage (edit: on same gpu architecture) only because its wider and is exactly what xsx vs ps5 showing (bandwidth can be important but you can have narrow with high bandwidth also so its different topic)

 

[iroboto]

Its impossible to scale above teoretical tflops advantage only because its wider and is exactly what xsx vs ps5 showing (bandwidth can be important but you can have narrow with high bandwidth also so its different topic)

Yea. Right. The real issue there is we use TF as a singular number to represent a properly paired GPU, so that all aspects of the GPU work with the TF number. It’s generally worked for us, but lately this is skewing a bit off trend since Ampere.

But in the case of XSX it doesn’t work at all and we run into this problem where using TF is not a good metric for determining its performance.

 

[snc]

Yea. Right. The real issue there is we use TF as a singular number to represent a properly paired GPU, so that all aspects of the GPU work with the TF number. It’s generally worked for us, but lately this is skewing a bit off trend since Ampere.

But in the case of XSX it doesn’t work at all and we run into this problem where using TF is not a good metric for determining its performance.

when we have same arch works very good, xsx has this slight resolution advantage in many games despite so different api available for both consoles

 

[iroboto]

when we have same arch works very good, xsx has this slight resolution advantage in many games despite so different api available for both consoles

From what reviewers can measure I suppose. Drs makes this incredibly difficult to perceive. Ranges may not necessarily be the averages. Though perhaps one day soon they will have a way to actually measure performance through DRS to get a better idea of what’s happening instead of looking for frames to pixel count.

 

[Seanspeed]

From what reviewers can measure I suppose. Drs makes this incredibly difficult to perceive. Ranges may not necessarily be the averages. Though perhaps one day soon they will have a way to actually measure performance through DRS to get a better idea of what’s happening instead of looking for frames to pixel count.

Are you honestly trying to deny that XSX has a clear resolution advantage in numerous games? :/

 

[iroboto]

Are you honestly trying to deny that XSX has a clear resolution advantage in numerous games? :/

I go by data. Ranges are good. There has been no real data on DRS performance. I have my thoughts that Xbox is ahead in average because reported resolution ranges are higher.

that is not the same thing as an average resolution. And this particular metric is still yet to be discovered.

 

[PSman1700]

XSX should be more capable (GPU), but it doesnt show. Like Iroboto thinks, it may be due to its relatively 'odd' design to go slow for a RDNA2 product. Ampere/RDNA2/Turing all go wide & fast, relatively.

 

[iroboto]

XSX should be more capable (GPU), but it doesnt show. Like Iroboto thinks, it may be due to its relatively 'odd' design to go slow for a RDNA2 product. Ampere/RDNA2/Turing all go wide & fast, relatively.

Time will tell. It's really DRS and TAAU that is really muddying the waters for us to see. If we can get metrics on a per frame basis over a large enough data set, we can make better conclusions than we are now.
I want to be clear that before the consoles came out, we all made a lot of assumptions on a lot of incomplete data. It was fun debating performance with the limited information we had. It turns out, that I made a lot of assumptions that looking back I certainly not correct on a lot of things, but not completely wrong either. But the lesson is learned regardless. Consoles won't behave like PC parts necessarily, even if they could.

With the consoles now here in our hands, it's best to drop behind those pre-conceived notions of how we think they should perform, and just measure them as they are. There are pitfalls that we fall into if we keep looking back at our initial guesses at performance. I think this is an issue that NXG and a lot of people have a habit of falling into. It's habitual correlate the data with how you think the hardware works, instead of using the data to check other facets to really see if that behaviour is true. We need more data points, but this is largely why I like DF's work in this area more than others. They do a little bit more testing to double verify things they catch.

 

[Dr. Nick]

XSX should be more capable (GPU), but it doesnt show. Like Iroboto thinks, it may be due to its relatively 'odd' design to go slow for a RDNA2 product. Ampere/RDNA2/Turing all go wide & fast, relatively.

Could it also be that DXR is not as performant right now as opposed to using something that is more direct to the metal? In some cases, I think DirectX is getting in the way more than it should but that is just a guess.

 

[mr magoo]

The thing is we can discuss to no end wich design is better and wich console should perform better but the bottom line is both console design was influenced by the feedback from top dev studios from sony and msft. This is what they wanted and this is what they are comfortable with. We should closely look at what top dev studios from both sides are doing with hardware.

 

[dobwal]

Yea. Right. The real issue there is we use TF as a singular number to represent a properly paired GPU, so that all aspects of the GPU work with the TF number. It’s generally worked for us, but lately this is skewing a bit off trend since Ampere.

But in the case of XSX it doesn’t work at all and we run into this problem where using TF is not a good metric for determining its performance.

Can't say that. We know that the XSX is abnormally wide given its structure versus other RDNA2 cards but you can't parse out the hardware from the influence of the software and devs' abilities and comfort at this point. Xbox devs are supporting 4 pieces of hardware across 2 CPU archs, 2 GPU archs (with 2 different gens of GCN) and three unified different memory setups on the hardware as of now with ESRAM (one S), split performance (SXS) and traditional (one x).

Narrower designs have their own cons as the faster you go the greater the cycle latency between main memory and the compute unit.

Plus, the biggest monkey wrench for comparing TF across RDNA2 based platforms is AMD Infinity Cache, which makes going narrower but faster easier.

 

[Shortbread]

With Sony making strides to expand its marketplace with big yet delayed releases on PC, they do not come bigger than God of War on PC. Following my coverage of the PC port over on IGN, here I go deep on many technical and logistical reasons and choices of that port. Why DX11? A thorough and complete PC performance analysis across AMD and Nvdia, CPU usage, Input latency improvements, loading and more. What are you waiting for, hit that play button now!

 

[iroboto]

Can't say that. We know that the XSX is abnormally wide given its structure versus other RDNA2 cards but you can't parse out the hardware from the influence of the software and devs' abilities and comfort at this point. Xbox devs are supporting 4 pieces of hardware across 2 CPU archs, 2 GPU archs (with 2 different gens of GCN) and three unified different memory setups on the hardware as of now with ESRAM (one S), split performance (SXS) and traditional (one x).

Narrower designs have their own cons as the faster you go the greater the cycle latency between main memory and the compute unit.

Plus, the biggest monkey wrench for comparing TF across RDNA2 based platforms is AMD Infinity Cache, which makes going narrower but faster easier.

Sure, 'at all' is too large a stretch I agree. But the GPU is largely setup like an awkward 5700XT. With less ROPS. But with more DCUs. It's not a typical configuration we would normally see out there.

I suppose the question I should ask is, what defines a GPU as being narrow wrt RDNA2? Is it the number of shader engines? or the number of DCUs?
Because if we look at terminology of narrow and wide if a 6800xt is considered wide and a 6700 is considered narrow. Looking at them, you can see they have the same components, the only difference being that the 6800XT has 4 shader engines vs 6700 2 shader engines. But the composition of those shader engines are identical.

The only reason I make this distinction is because the shader engine contains all the fixed function items to support the compute units. Thus, increasing the number of shader engines increases the whole pipeline and performance together, where increasing just Compute units does not.

 

[dobwal]

Sure, 'at all' is too large a stretch I agree. But the GPU is largely setup like an awkward 5700XT. With less ROPS. But with more DCUs. It's not a typical configuration we would normally see out there.

I suppose the question I should ask is, what defines a GPU as being narrow wrt RDNA2? Is it the number of shader engines? or the number of DCUs?
Because if we look at terminology of narrow and wide if a 6800xt is considered wide and a 6700 is considered narrow. Looking at them, you can see they have the same components, the only difference being that the 6800XT has 4 shader engines vs 6700 2 shader engines. But the composition of those shader engines are identical.

The only reason I make this distinction is because the shader engine contains all the fixed function items to support the compute units. Thus, increasing the number of shader engines increases the whole pipeline and performance together, where increasing just Compute units does not.

Maybe MS de-emphasized ROPs in their design because their profiling didn't identify ROPs as a common bottleneck. I mean if you are developing titles on UE5 or any engine that can heavily employ compute based rasterizers, how many more ROPs do you need this gen?

MS probably had a more stringent transistor budget than AMD PC gpus. The general trend has been moving more and more work off the fixed function side onto the compute shader side. If you are MS it might more sense to balance your design to be more compute based heavy versus obliging a compute to fixed function ratio thats a product of PC design.

AMD and Nvidia are hamstrung by the fact that performance of new PC designs are mostly measured by old software. Consoles are not.

 

[iroboto]

AMD and Nvidia are hamstrung by the fact that performance of new PC designs are mostly measured by old software. Consoles are not.

elegant sentence.

 

[PSman1700]

MS probably had a more stringent transistor budget than AMD PC gpus. The general trend has been moving more and more work off the fixed function side onto the compute shader side. If you are MS it might more sense to balance your design to be more compute based heavy versus obliging a compute to fixed function ratio thats a product of PC design.

AMD and Nvidia are hamstrung by the fact that performance of new PC designs are mostly measured by old software. Consoles are not.

Yet the PC gpu's are both wide and fast.

 

[pjbliverpool]

AMD and Nvidia are hamstrung by the fact that performance of new PC designs are mostly measured by old software. Consoles are not.

They are for at least the first year or two though which is why we're talking about this. If the theory is correct then the XSX's more future looking architecture is currently underperforming because its being measured on 'old' cross gen software. And it's the early impressions which can set the sales course for the entire life of a console.

It's also worth considering the importance of backwards compatibility and remasters which are more important than ever to the console market and would likely depend more on the older software paradigm.

 

[iroboto]

Yet the PC gpu's are both wide and fast.

each generation of PC GPU has only gotten faster. And none have ever opted to leverage a node shrink to only increase clock speed when there was clearly an opportunity to increase silicon realestate as well.
Which is why when the first terms of fast and narrow and wide and slow came up, I did pull back on this on several occasions.
PS5 doesn't do anything special from the architecture of a 5700XT/6700. It's really the XSX that is out of composition here with respect to its relative DCU / FF hardware ratios. Fixed clock speeds are the norm for consoles; they are a cost saving measure to improve yields for lower grade silicon. The reason it's slow it's because it's cheap, not because XSX is particularly wide.

MS could have gone the variable clock speed route, but they would have likely blown the price budget.

 

[PSman1700]

the XSX's more future looking architecture

You'd wonder what is so 'future looking' there, i mean, its basically RDNA2 what we have on PC, a 6800 is just as futuristic i think. If not more so.

each generation of PC GPU has only gotten faster. And none have ever opted to leverage a node shrink to only increase clock speed when there was clearly an opportunity to increase silicon realestate as well.
Which is why when the first terms of fast and narrow and wide and slow came up, I did pull back on this on several occasions.
PS5 doesn't do anything special from the architecture of a 5700XT/6700. It's really the XSX that is out of composition here with respect to its relative DCU / FF hardware ratios. Fixed clock speeds are the norm for consoles; they are a cost saving measure to improve yields for lower grade silicon. The reason it's slow it's because it's cheap, not because XSX is particularly wide.

MS could have gone the variable clock speed route, but they would have likely blown the price budget.

You sure? Looking at Turing and subsequently Ampere seemingly have gotten wider and wider, atleast in NV's case the clocks havent boosted all that much. For AMD yes, in special RDNA2 we are in the very high clock range. Also on console in PS5's case, for a console 2.26ghz is very high.
On paper the XSX's GPU would be 'better specced', as was speculated in these forums. Yet the PS5 gpu is doing atleast as well/good in the games we can test them against. I think the most intresting bechmarks would be to compare vs PC gpus and see where the XSX would slot (same for the PS5). PS5 vs XSX is much more complicated to test i think.

In some games like Control the XSX GPU is almost doing what the paper specs say though (18% or so faster), which is intresting.