Digital Foundry Article Technical Discussion [2021]

Status
Not open for further replies.
How can you possibly say that? It's still a relatively new architecture and is having to battle methods built from a mature RTX implementation.

RTX saw some pretty substantial gains from both driver and game level optimizations as developers learned the hardware so I expect RDNA2 to show some improvement.

Insomniac are showing with Ratchet and Clank that RDNA2 isn't actually that bad an RT.

Would it make you feel better if i said in comparison to pc parts of today. Think about it ps4 got the pro what 3 years in and the one x hit 4 years after their respective launch. What would that be 2023/2024. We could be comparing them to RDNA 4 and Geforce 5x00 series in terms of ray tracing. You think a faster clocked higher core rdna 2 in a playstation 4 pro or xbox series x ? I'd argue the consoles would fall even further behind
 
How can you possibly say that? It's still a relatively new architecture and is having to battle methods built from a mature RTX implementation.

RTX saw some pretty substantial gains from both driver and game level optimizations as developers learned the hardware so I expect RDNA2 to show some improvement.

Insomniac are showing with Ratchet and Clank that RDNA2 isn't actually that bad an RT.
We do know the relative performance between RTX and RDNA in RT in the PC space, and in general, RDNA is lacking for sure. A lot of the optimizations that happened to PC titles on RTX simply involved doing less work. They were halving their ray count because they thought RTX would be faster than it turned out to be. But that doesn't change the fact that RDNA is even slower for RT.

RDNA has a cost for RT beyond it's theoretical limits anyway. Because they are using compute units for RT, unlike RTX that has dedicated RT cores, that hardware can't be used for the compute needed to eventually render the scene. The RT acceleration in current hardware is mostly centered around figuring out what needs to be rendered, so if the final scene has a higher compute cost (more stuff is being rendered in those reflections, and all that needs to be shaded as well) and you are spending your compute on ray intersections instead of rendering you are going to be at a performance disadvantage when compared to hardware that has dedicated hardware for each task. Assuming there are similar levels of compute like there are in RDNA and RTX like there is now.
 
What I think Insomniac are showing is that Insomniac are crazy good developers.
Well they are also showing something very specific, good quality concessions based on their art. They have it so there are no cubemaps/secondary reflections in their reflections. They have it so no transparency particles are in reflections. They use a lower LOD of geometry in their reflections. In their 4K mode, they use checkerboard RT Reflections. But they note that the checkerboarded 4K RT Reflection only shoots out 20% more rays or so than quarter resolution ray traced reflections at that same res: they instead rely on temporal information more heavily. Trading cleanliness in motion for a greater resolve with less motion.

Similarly their 60 fps mode makes other smart concessions regarding objects in the BVH, shadows in the BVH, and the amount of rays shot out again (quarter res).

Basically what I am trying to say is that they make smart concessions - that is the mark of great single platform development. I do not think the reason their RT implementation is so successful because they have discovered some secret way to enhance performance or quality due to deft to-the-metal coding - rather they know what to cut when and where. Completely different than a developer just using the default UE4 reflections at full resolution and having poor performance or image quality.
 
What I think Insomniac are showing is that Insomniac are crazy good developers.
Their true claim to fame is never missing a date while juggling multiple projects with a high bar for quality. I can’t think of any other developer that hits their targets so well. Maybe playground games, we will see how they do juggling fable with forza.

Incredible management and process and tools pipeline. They are exceptionally good at what they do.

it doesn’t mean that they could deliver a proper cyberpunk; but that they are very good at understanding their delivery potential within a target timeframe and maximizing it.
 
Well they are also showing something very specific, good quality concessions based on their art. They have it so there are no cubemaps/secondary reflections in their reflections. They have it so no transparency particles are in reflections. They use a lower LOD of geometry in their reflections. In their 4K mode, they use checkerboard RT Reflections. But they note that the checkerboarded 4K RT Reflection only shoots out 20% more rays or so than quarter resolution ray traced reflections at that same res: they instead rely on temporal information more heavily. Trading cleanliness in motion for a greater resolve with less motion.

Similarly their 60 fps mode makes other smart concessions regarding objects in the BVH, shadows in the BVH, and the amount of rays shot out again (quarter res).

Basically what I am trying to say is that they make smart concessions - that is the mark of great single platform development. I do not think the reason their RT implementation is so successful because they have discovered some secret way to enhance performance or quality due to deft to-the-metal coding - rather they know what to cut when and where. Completely different than a developer just using the default UE4 reflections at full resolution and having poor performance or image quality.
Which console games have full resolution RT running at 60fps at 1440p on average? Not Watchdog Legion with quarter resolution RT at only 30fps compared to the higher res (double?) RT and 60fps for Insomniac titles (and visible at a much longer distance). Compared to Watchdogs Legion RT (or any game having RT on console) it seems they have found some advantage coding to the metal whether you like it or not because the difference to most people is staggering.

Remember, it's RT, 60fps at mostly ~1440p on Insomniac games, you can't do that only thanks to smoke and mirrors. The others games can only do about the same final result at half the framerate.
 
Which console games have full resolution RT running at 60fps at 1440p on average? Not Watchdog Legion with quarter resolution RT at only 30fps compared to the higher res (double?) RT and 60fps for Insomniac titles (and visible at a much longer distance). Compared to Watchdogs Legion RT (or any game having RT on console) it seems they have found some advantage coding to the metal whether you like it or not because the difference to most people is staggering.

Remember, it's RT, 60fps at mostly ~1440p on Insomniac games, you can't do that only thanks to smoke and mirrors. The others games can only do about the same final result at half the framerate.
How many games get rid of cubemaps in their RT reflections (WDL does not, same with control)? How many games use a completely different geo LOD in their reflections (no other game using DX12/Vulkan native/or on console does this atm)? They are obviously cutting corners in areas that have meaningful performance implications.

Also Ratchet at 60 fps is not just "~1440p". It is pretty often 1296p or even lower when the action starts or if a reflection is full screen (like the dancing scene for example). The whole point of DRS is to increase res as much as possible given GPU load - if a game was constantly just around 1440p that would mean every view and area in the game has the exact same GPU performance cost which is absurd.

And to just use a platitude of sorts: there is evidence of something (obvious quality cuts), but we have zero outward evidence of other things (hyper to the metal coding that magically beats every other game programmer out there). I prefer to use the thing we have evidence of as a basis for reasoning rather than a belief.
 
Last edited:
How many games get rid of cubemaps in their RT reflections (WDL does not, same with control)? How many games use a completely different geo LOD in their reflections (no other game using DX12/Vulkan native/or on console does this atm)? They are obviously cutting corners in areas that have meaningful performance implications.

Also Ratchet at 60 fps is not just "~1440p". It is pretty often 1296p or even lower when the action starts or if a reflection is full screen (like the dancing scene for example). The whole point of DRS is to increase res as much as possible given GPU load - if a game was constantly just around 1440p that would mean every view and area in the game has the exact same GPU performance cost which is absurd.

And to just use a platitude of sorts: there is evidence of something (obvious quality cuts), but we have zero outward evidence of other things (hyper to the metal coding that magically beats every other game programmer out there). I prefer to use the thing we have evidence of as a basis for reasoning rather than a belief.
generaly agree tough they later significantly improved rt resolution in ratchet not bringing new tricks but rather finding some algorithm optimization
 
Last edited:
generaly agree tough they later significantly improved rt resolution in ratchet not bringing new tricks but rather finding some algorithm optimization

Or in Frosbite video, they said for Surfel GI they need to use intrinsics and optimize per platform. Devs can always optimize for a platform.
 
Or in Frosbite video, they said for Surfel GI they need to use intrinsics and optimize per platform. Devs can always optimize for a platform.

Insomniac has only ever had to code for the ps4 / ps5 so it should be easier when all your resources pour into one or two platforms and spiderman engine was already optimized a lot for ps4 .

Same with the team behind Ratchet , they only had to target ps5. While some of the other games talked about are targeting last gen , this gen , pc and maybe switch
 
Insomniac has only ever had to code for the ps4 / ps5 so it should be easier when all your resources pour into one or two platforms and spiderman engine was already optimized a lot for ps4 .

Same with the team behind Ratchet , they only had to target ps5. While some of the other games talked about are targeting last gen , this gen , pc and maybe switch
Insomniac shipped at least 2 games for Xbox One, and only 3 or 4 for PS4 (depending on if you count Miles as it's own game and not an expansion). Also 7 games on PC since the launch of the PS4.

Edit - As soon as I hit post I occurred to me that you are specifically talking about how optimized their exclusive games are compared to other multiplatform titles and not making a statement about them never making games for other platforms. I'll just leave my original post unchanged above so people can see my foolishness and lack of understanding of the context.
 
Last edited:
It's hard to keep track of who says what over the years, but it seems like depending on the argument, it always flip flops back and forth between exclusive developers not achieving more due to single platform development and having a huge benefit because of single platform development.
 
Insomniac shipped at least 2 games for Xbox One, and only 3 or 4 for PS4 (depending on if you count Miles as it's own game and not an expansion). Also 7 games on PC since the launch of the PS4.

Edit - As soon as I hit post I occurred to me that you are specifically talking about how optimized their exclusive games are compared to other multiplatform titles and not making a statement about them never making games for other platforms. I'll just leave my original post unchanged above so people can see my foolishness and lack of understanding of the context.

Lol yea that is what I was talking about.

Obviously if you have say a 3 year development time frame and you have to develop for 1 or 2 systems vs 4 or 5 your going to be able to make the game run much better on just those 1-2 systems vs 4 or 5.

Esp when you go from the switch all the way up to the next gen systems . Add in pc which itself is a huge beast with hundreds of configurations you have to deal with. Well its obvious your going to have a more performant game on that 1-2 system release.

So for spiderman well the first game got a ton of optimization since it was only for the playstation 4. They used the same engine for miles so I am sure there was not to much to do for the ps4 version and that lead to having a lot of time to optimize for the ps5 and ray tracing and again they shipped the game and continued to work on boosting just the ps5 performance and ray tracing performance.
 
So oddly, GN results came closer to what I thought the end results would be; increase of 5 degree exhaust likely just meant 5 degrees at the SoC for what they measured.
I guess I'm pretty happy with that - the formula worked lol, though different measurements will likely come up with different things. But neat, seems like Sony succeeded in largely creating a better heatsink with less costs.
 
Last edited:
So oddly, GN results came closer to what I thought the end results would be; increase of 5 degree exhaust likely just meant 5 degrees at the SoC for what they measured.
I guess I'm pretty happy with that - the formula worked lol, though different measurements will likely come up with different things. But neat, seems like Sony succeeded in largely creating a better heatsink with less costs.
Shocking that Sony engineers know how hat they are doing? ;)
 
Status
Not open for further replies.
Back
Top