Support for Machine Learning (ML) on PS5 and Series X?

Yes, and to be clear I do not think this case of "PS5 running ML inference" serves as any proof whatsoever on whether or not the PS5's ALUs can run dot4 INT8 / dot8 INT4 operations.
For all I know, this NN could be so light that it's running on a small percentage of one CPU core, and running it on the GPU could just bring unnecessary dev time with little to no performance benefit.
yes.
I can see the CPU being freed to do this type of work especially if you're doing animation on the CPU side of things.
 
Well it would appear it is indeed realtime

B2ZTNUi.jpg
 
also no performance penalty
All the 9th-gens are really green so we should still expect a ramping up of better effects with no performance penalty and/or better performance at similar IQ.

In this particular case, it seems Insomniac is keen on using Miles Morales to learn about the PS5's capabilities, and I wouldn't be surprised if Sony's ICE team is using the game to deploy prototypes for several technologies.

Miles Morales was a day-one PS5 title that released with a 60FPS raster-only mode and a 30FPS raytracing mode. A couple of months later they patched it to get a 60 FPS + raytracing mode, and now they're releasing another patch for muscle deformation using ML inference.


yes, looks like msft cant do anything right. Maybe for the next gen they will hire real engineers
This is unnecessary baiting...
 
  • Like
Reactions: snc
Also a bit ironic that Microsoft screams about ML capabilities at every chance while Sony stays silent on that theme, yet apparently the first ML inference implementation being deployed in an actually released game is coming from a Sony 1st-party title.

I will repeat it again Sony has a recontruction/AA deep lenarning based method available but not used in any game for the moment. We need to be patient maybe Horizon Forbidden West or 2022 first and third party games.
 
Last edited:
  • Like
Reactions: snc
Could they use the Tempest power available to devs to do this? For multiplat devs that could be a good way to replace the (supposedly) missing int4 / int8 hardware on PS5.
doubt, tempest is very capable for 3d audio but not a compute monster ;d
 
Also a bit ironic that Microsoft screams about ML capabilities at every chance while Sony stays silent on that theme, yet apparently the first ML inference implementation being deployed in an actually released game is coming from a Sony 1st-party title.
I believe the situation is similar to the launch multiplatform games where people were surprised with the results that PS5 could keep up with Xbox and celebrated how big of genius Cerny is. But only recently we started to see differences in the multiplatform games and goalposts were moved. Just like with SSD.

I do wonder what benefits the dedicated ML hardware will bring to Xbox games. MS just doesn't have the next gen game for now to show off.
 
I believe the situation is similar to the launch multiplatform games where people were surprised with the results that PS5 could keep up with Xbox and celebrated how big of genius Cerny is. But only recently we started to see differences in the multiplatform games and goalposts were moved. Just like with SSD.

I do wonder what benefits the dedicated ML hardware will bring to Xbox games. MS just doesn't have the next gen game for now to show off.
ML is largely supported by MS, as in, they support the community by providing environments and tools for deployments. But generally speaking, outside of that space they so far have a limited products that they sell to companies to leverage.

I think the model you're seeing here (third party ML teams working with studios) is likely to be the future of ML support in gaming. Studios specialize particularly in developing and deploying content, audio, graphics, animation etc. To tack onto a studio, a ML R&D team is significant, and outside of Epic and Unity whose main goal is engine development, it's hard to see small studios pick this up. The engineers just want a model ultimately to integrate into the game, but they don't want the overhead of costs of creating those models (creating training sets, paying data scientists). So it makes sense for studios to have this type of relationship, if you want to see ML happen on games in general, there has to be a good fit for what a third party studio is offering and what a gaming studio requires.

It is fantastic and warming to my heart to see ML get deployed into games that have nothing to do with Super resolution. I think this is where ML deployments are going to be the most exciting to watch.
 
I will repeat it again Sony has a recontruction/AA deep lenarning based method available but not used in any game for the moment. We need to be patient maybe Horizon Forbidden West or 2022 first and third party games.
Where did you see info on that? Patents?

My current guess is both consoles will just use FSR, which doesn't seem to be using ML at all.


Could they use the Tempest power available to devs to do this? For multiplat devs that could be a good way to replace the (supposedly) missing int4 / int8 hardware on PS5.
I don't think Tempest is good for many things other than audio, nor would it represent any significant boost in compute performance (1 CU @ 2.23GHz is just ~0.57 TFLOPs FP16 / 1.140 TOPs INT8). It's a single CU without caches and ML inference ops probably need those. Also, it looks so different from a RDNA2 "half-WGP" that apparently no one can tell exactly where it is, from the SoC's x-ray pictures.




I believe the situation is similar to the launch multiplatform games where people were surprised with the results that PS5 could keep up with Xbox and celebrated how big of genius Cerny is. But only recently we started to see differences in the multiplatform games and goalposts were moved. Just like with SSD.
Yes, of course. All these newly-formed fans of Photo Mode really taught Cerny and Sony's hardware teams a thing or two about real-time videogame render performance.
And apparently we're at the point of making up issues with SSD performance too. Neat.
 
Where did you see info on that? Patents?

My current guess is both consoles will just use FSR, which doesn't seem to be using ML at all.



I don't think Tempest is good for many things other than audio, nor would it represent any significant boost in compute performance (1 CU @ 2.23GHz is just ~0.57 TFLOPs FP16 / 1.140 TOPs INT8). It's a single CU without caches and ML inference ops probably need those. Also, it looks so different from a RDNA2 "half-WGP" that apparently no one can tell exactly where it is, from the SoC's x-ray pictures.





Yes, of course. All these newly-formed fans of Photo Mode really taught Cerny and Sony's hardware teams a thing or two about real-time videogame render performance.
And apparently we're at the point of making up issues with SSD performance too. Neat.

From a dev I know . This has nothing to do with Super Resolution and AMD initiative, this is something else developed by Sony for PS5.
 
Yes of course. But what I am trying to understand is the part where the Insomniac dev claims their real-time ML inference does not change current game performance.

Maybe it is using double rate FP16, this is available since PS4 PRo. ML inference not only use INT 8 or INT4.
 
Back
Top