Digital Foundry Article Technical Discussion [2022]

[Globalisateur]

To answer to the critique put forward here: we have heard from devs making games which target heavy CPU usage that the GPU in PS5 does in fact downclock. But whether an end user notices this in a game with TAA, post-processing and DRS is a whole other question. That is the point of the PS5 Design.

For example - think of a game with an unlocked framerate to 120 with DRS targetting a high output res. How exactly does that fit into a fixed and shared power budget? The obvious answer is it stresses both CPU and GPU to their max and power adjusts

Of course it downclocks. Cerny told us that 2 years ago! But does it noticeably impact the final performance? the answer is in the vast majority of time, no. This is what we have seen in many open world games that performs as well or even better than expected. In the end the PS5 is behaving like a 10TF machine (even better in many cases) even in those demanding game and the dynamic clocks system does not impact on actual games performance. It's not only because of DRS or TAA; Cerny patiently explained it in several interviews:

- The downclocks can be very short (a few ms) so that means on average the drop will be meaningless on the whole 16ms frame when they occur.
- Some drops will have no impact on performance as the system (CPU or GPU) can actually downclock if it reckons it won't be used (idle) during this time, as explained by Cerny:

There's another phenomenon here, which is called 'race to idle'. Let's imagine we are running at 30Hz, and we're using 28 milliseconds out of our 33 millisecond budget, so the GPU is idle for five milliseconds. The power control logic will detect that low power is being consumed - after all, the GPU is not doing much for that five milliseconds - and conclude that the frequency should be increased. But that's a pointless bump in frequency...At this point, the clocks may be faster, but the GPU has no work to do. Any frequency bump is totally pointless. "The net result is that the GPU doesn't do any more work, instead it processes its assigned work more quickly and then is idle for longer, just waiting for v-sync or the like. We use 'race to idle' to describe this pointless increase in a GPU's frequency...So, when I made the statement that the GPU will spend most of its time at or near its top frequency, that is with 'race to idle' taken out of the equation...The same is true for the CPU

https://www.eurogamer.net/articles/digitalfoundry-2020-playstation-5-the-mark-cerny-tech-deep-dive

But whether an end user notices this in a game with TAA, post-processing and DRS is a whole other question.

We don't need to make such mysterious questions. Have you recently noticed the PS5 having dropped frames or resolution during demanding scenes? For instance using the currently most demanding (and next-gen) benchmark, the RT / compute heavy UE5 demo. Do you think the PS5 behaves like a 9 Tf machine compared to a 12TF machine here?

 

[dobwal]

Running ML/AI on seperate hw cores means less impact on the gpu shaders itself, like with ray tracing.

And? The question is whether the work performed by the gpu shaders for upscaling + work performed by shaders rendering at a sub-targeted resolution is appreciably less than the shader work needed for native rendering at the target resolution.

If a gpu can render a 1080p frame at 12ms and use another 4ms to produce a quality upscale to 4K, while rendering at a native 4K takes 33.3ms, then its a win regardless. The impact of the upscale on shaders is beneficial because the overall cost of producing a 4k frame is reduced by half.

Having dedicated hardware would be better as the gpu shaders don't have to participate but it doesn't mean doing all the work on the shaders is not a viable solution. Its just a matter of how fast the shaders can perform the work at an acceptable quality.

 

[dobwal]

"Die on that hill" / "The hill you want to die on" is an English idiom and nothing more.

https://grammarist.com/idiom/the-hill-you-want-to-die-on/

The hill you want to die on stems from 20th-century American literary works related to military origins. It is often used in a questioning form to ask if an opinion or action is truly worth the effort.

It also can be used to strengthen an argument further; that something is important enough to die upon that hill. In this case, the hill is meant to represent a struggle worth fighting for.​

I just remember a bunch of war movies like Hamburger Hill. Elevated positions have a strategic advantage on a battlefield so areas like hills were often fortified as defensive positions. During the Vietnam war, the Viet Cong would often take position on a hill, bait Americans into a battle, dish out losses while Americans tried to take the hill and then flee to avoid heavy losses. Since the hills often had no strategic value, the American military would abandoned them and the Viet Cong would redeploy for a rinse and repeat strategy. A lot of american lives were lost on these hill battles.

 

[dobwal]

Of course it downclocks. Cerny told us that 2 years ago! But does it noticeably impact the final performance? the answer is in the vast majority of time, no. This is what we have seen in many open world games that performs as well or even better than expected. In the end the PS5 is behaving like a 10TF machine (even better in many cases) even in those demanding game and the dynamic clocks system does not impact on actual games performance. It's not only because of DRS or TAA; Cerny patiently explained it in several interviews:

- The downclocks can be very short (a few ms) so that means on average the drop will be meaningless on the whole 16ms frame when they occur.
- Some drops will have no impact on performance as the system (CPU or GPU) can actually downclock if it reckons it won't be used (idle) during this time, as explained by Cerny:


https://www.eurogamer.net/articles/digitalfoundry-2020-playstation-5-the-mark-cerny-tech-deep-dive

We don't need to make such mysterious questions. Have you recently noticed the PS5 having dropped frames or resolution during demanding scenes? For instance using the currently most demanding (and next-gen) benchmark, the RT / compute heavy UE5 demo. Do you think the PS5 behaves like a 9 Tf machine compared to a 12TF machine here?

Unless you are monitoring the clock frequency you would haves no ideal of when or where the PS5 is downclocking. And most quality devs are going to optimize to avoid downclocking that actually impacts gameplay. You will never see a PS5 version of PS4 God of War because the PS5 won't actually allow it. You can't have a circumstance where the PS5 sounds likes its killing its cooling fan as it maintains its max clocks under heavy workloads like a PS4 does for demanding games.

Whats the difference between a 9TF vs 10TF PS5? None of us have any practical ideal as its too small a gap to readily determine how the difference would affect the performance of a title.

 

[see colon]

Worse, they go on to conclude that because, in these scenarios the PC loses performance from having vsync on, the PS5 which is also running with vsync, but completely locked as 60fps, is losing a similar amount of performance.

While this would be true in a general sense, there are games that, for one reason or another, run worse with the in game settings for vsync on than they do when it is forced on, or framerate limited at the driver level.

Odd behavior like this is also not limited to PC. There were games, and I don't remember titles right now, but there was at least one that ran worse or presented worse when displayed at 120hz. It was a game that didn't support 120fps, I think it was an Xbox One title. And then stuff like Red Dead 2 on PS4 pro, where you would get a blurrier image when your output is set to 4k than it was when set to 1080p. I just don't think that doing comparisons where you intentionally show a game configured in a way that gives you a worse experience is a great way to do it.

 

[Deleted member 11852]

Exactly. What we've seen in the past from other Youtubers are comparisons between the PC and PS5 with the PS5 locked at 60fps and thus not experiencing this issue while the slightely less powerful PC is hitting this issue on account of occasionally missing the 16ms window. That's fine until said Youtubers start to draw percentage performance comparisons from those results. Essentially ignoring the vsync performance deficit.

There may be a glorious future where most gams ship with unlocked framerate VRR modes and settable resolutions (on console) which will show what each game/engine/platform can do but until then, this is continue to be an issue. Counting dropped frames is easy, estimating unused graphics/memory/bandwidth/CPU headroom isn't from observation, only devtools.

And there has been a fairly consistent weirdness on current generation consoles to skew same games on one console to aim for higher resolutions and sometimes drop a frame vs the other console where will invariably never drop a frame but run at slightly reduced frame rates. Which demonstrated even on those platforms, unexplainable differences exist and it's not clear if this is some actual platform difference or just different decisions made by particular platform teams in multi-platform projects. ¯\_(ツ)_/¯

 

[Arwin]

My guess would be that on one platform it is sometimes easier to get higher performance thanks to more TFs, but on the other platform it is easier to be stable, thanks to more predictable behavior.

 

[Shortbread]

It's been seven years since the launch of the original Dying Light on PC, Xbox One and PlayStation 4... and here's Techland delivering triple-A fan service with a comprehensive upgrade that doesn't just improve the game for Xbox Series X/S and PlayStation 5 but also adds some love for Xbox One X too. Oliver Mackenzie has the detailS.

 

[Riddlewire]

As I already stated in the Dying Light thread, the 1080p mode for Series S/One X is completely pointless. It has no reason to exist.
Better to unlock the framerate and allow VRR owners to get as much performance as possible.
Or drop the resolution to 900p and try to hit 60fps.

 

[iroboto]

My guess would be that on one platform it is sometimes easier to get higher performance thanks to more TFs, but on the other platform it is easier to be stable, thanks to more predictable behavior.

I mean, honestly something often not discussed but probably overlooked is that once developers hit a reasonable parity, there's little to no monetary return on investment to optimize further.

Perhaps this will be an issue for them later on when they continue to try to push the graphics envelope further.

 

[Pete]

As I already stated in the Dying Light thread, the 1080p mode for Series S/One X is completely pointless. It has no reason to exist.
Better to unlock the framerate and allow VRR owners to get as much performance as possible.
Or drop the resolution to 900p and try to hit 60fps.

The 1080p60 mode can be interesting as a technical comparison point to the 1440p60 mode just in terms of power draw. Hopefully AJ Gaming will take a look at that. Since fan noise isn’t an issue with the current gen, though, 1080p60 doesn’t seem interesting from a player’s perspective except mayyybe if you’re gaming in an uncomfortably hot room, in which case I’m not sure the rough equivalent of turning off a 40-60W bulb will make a real difference.

 

[Riddlewire]

The 1080p60 mode can be interesting as a technical comparison point to the 1440p60 mode just in terms of power draw. Hopefully AJ Gaming will take a look at that. Since fan noise isn’t an issue with the current gen, though, 1080p60 doesn’t seem interesting from a player’s perspective except mayyybe if you’re gaming in an uncomfortably hot room, in which case I’m not sure the rough equivalent of turning off a 40-60W bulb will make a real difference.

I'm not sure what you're talking about.
I was talking about the One X and Series S.
It's not 1080p 60fps on those. It's only 30fps.
The same framerate as the 1440p version. It's a meaningless mode on those machines.

 

[Pete]

I'm not sure what you're talking about.
I was talking about the One X and Series S.
It's not 1080p 60fps on those. It's only 30fps.
The same framerate as the 1440p version. It's a meaningless mode on those machines.

Whoops. I thought I saw “One X” running at 60fps in the video. Still, the 1080p30/1440p30 choice on X1X/XSS is as meaningless as the 1080p60/1440p60 choice on the PS5/XSX.

All those systems are supposed to be pretty quiet, too. It might make more sense on a PS4 Pro because its fan noise ramps with system utilization. (I play NMS at 1080p60 just because the fan noise at 4k30 is intolerable.)

 

[see colon]

As I already stated in the Dying Light thread, the 1080p mode for Series S/One X is completely pointless. It has no reason to exist.
Better to unlock the framerate and allow VRR owners to get as much performance as possible.
Or drop the resolution to 900p and try to hit 60fps.

My thoughts exactly. Even if they had to drop to 720p to hit 60fps, there would at least be a reason for the mode to exist.

 

[zed]

Best DF video ever made. No lie.

yes I actually watched this video, the others I skip through

 

[Dr. Nick]

There is a great deal of XeSS talk in here. I think the DirectML question should have been phrased as MLSS, which I believe is what most people thought when Microsoft showed off ML-based image reconstruction all of those years ago. Based on past Intel Integrated GPUs, I guess that XeSS dp4a should provide a good performance boost on even those chips, so I think it should run well on the Series S. The question will be just how good will it be? Could it be used to get a form of 1440p 60fps output consistently? Like most of this generation's features, my main concern is how well it will help boost the Series S performance.

If it is performant enough, since XeSS will be open source, I think if Microsoft had any plans of creating their own ML-based image reconstruction, they should shelve it and maybe focus on what they could do to improve it. I think when everything is said and done, all 3 GPU manufacturers will have their own ML-based reconstruction (FSR 3.0?). If two of these companies are making their version of it open source, I'm not sure if it will be worth it for Microsoft to continue to pursue it...unless they have something earth-shattering in the works.

P.S. I know someone will bring up the fact that the FSR currently is not using ML, but I think what we see from AMD is a slow progression towards using ML. They aren't ready to go all the way but I'm betting FSR 3 or 4 will use ML. How they are going to go about it is anyone's guess, but I predict they will eventually end up using dp4a before finally using something like Tensor/XMX cores.

 

[iroboto]

There is a great deal of XeSS talk in here. I think the DirectML question should have been phrased as MLSS, which I believe is what most people thought when Microsoft showed off ML-based image reconstruction all of those years ago. Based on past Intel Integrated GPUs, I guess that XeSS dp4a should provide a good performance boost on even those chips, so I think it should run well on the Series S. The question will be just how good will it be? Could it be used to get a form of 1440p 60fps output consistently? Like most of this generation's features, my main concern is how well it will help boost the Series S performance.

If it is performant enough, since XeSS will be open source, I think if Microsoft had any plans of creating their own ML-based image reconstruction, they should shelve it and maybe focus on what they could do to improve it. I think when everything is said and done, all 3 GPU manufacturers will have their own ML-based reconstruction (FSR 3.0?). If two of these companies are making their version of it open source, I'm not sure if it will be worth it for Microsoft to continue to pursue it...unless they have something earth-shattering in the works.

P.S. I know someone will bring up the fact that the FSR currently is not using ML, but I think what we see from AMD is a slow progression towards using ML. They aren't ready to go all the way but I'm betting FSR 3 or 4 will use ML. How they are going to go about it is anyone's guess, but I predict they will eventually end up using dp4a before finally using something like Tensor/XMX cores.

my thoughts here: it's in MS favour to make their own as it's unlikely intel or nvidia will give the model away for free for alteration. But MS could allow modification in theory as long as it's used in conjunction with directX as a stipulation. Which would be a generous give away to companies looking to expand on existing techniques. Even if MS didn't give it away, I think having a built in MLSS with DirectX as a bolt on into the pipeline would be a straight up win for them.

 

[PSman1700]

I predict they will eventually end up using dp4a before finally using something like Tensor/XMX cores.

RDNA3 is bound this year i think? RDNA4 could be it next year or there-after.
Thanks DF for explaining the stuff around ML/AI reconstruction as to why AMD gpus dont have it (lack of hw acceleration).

 

[davis.anthony]

Like most of this generation's features, my main concern is how well it will help boost the Series S performance.

I wonder if XSS has enough performance to actually use it without killing frame times, XSX has half the INT4 TOPS as an RTX2060 and XSS's GPU is 1/3 of XSX's.

XSS has 1/6? (My maths is fuzzy) the INT4 TOPS as an RTX 2060, so is that even enough to do an ML based upscale in a reasonable amount of frame time?

 

[techuse]

I wonder if XSS has enough performance to actually use it without killing frame times, XSX has half the INT4 TOPS as an RTX2060 and XSS's GPU is 1/3 of XSX's.

XSS has 1/6? (My maths is fuzzy) the INT4 TOPS as an RTX 2060, so is that even enough to do an ML based upscale in a reasonable amount of frame time?

AFAIK we don't have any info on whether or not matrix math throughput is the limiting factor in DLSS performance.