AMD FidelityFX on Consoles
- Thread starter invictis
- Start date
I thought that neural network was done outside of the PS5, and then the results of the muscle information were implemented in the game.
Machine learning inference is done real-time on PS5
Perhaps you're thinking of the training component, which is done offline on supercomputers?
I don't get why anyone would think that any device capable of performing an heuristic search of a finite space to find the most optimal solution isn't capable of ML. Because it is, it just boils down to how fast that search is and how big that space can be. And you don't need supercomputers to train with, you could use a Pi4. It would just take you a little bit longer 
D
Deleted member 13524
Guest
Neural Network Training: creating a system of equations that provide a desired output given an input.
Neural Network Inferencing: providing inputs to the pre-trained Neural Network (i.e. running the equations) to achieve the desired output.
Training happened outside the PS5. The trained network is basically what Ziva Dynamics sold to Sony / Insomniac, but the inferencing is happenning in the PS5 GPU.
come on -- this forum is so silly. Everyone here should understand what hardware acceleration vs software means. It's completely unknown how much of a hardware accelerated advantage (if any) the xsx has over the ps5 on machine learning, but that's what's being speculated about. You can do machine learning inference on a toaster cpu -- it's just a question of how fast, and how much it competes with other resources.
If (big if!) the xsx has some powerful hardware accelerated machine learning advantage, that'd be very significant for a dlss-like situation, where you need to do a lot of ML calculation but it defeats the purpose if you're competing with your own renderer for gpu time.
My thoughts; the performance difference will largely be governed by the performance of the model before performance of their respective hardware.
if they are lucky enough to use the same models then you can have a closer comparison on performance i think. But I’m not sure if IP restrictions around models will be a factor. Just thinking out loud. If MS or Sony deploys their own algorithms for upscaling or Aa; it will not show up on the competing platform.
I think I know why, most people have no ide what "an heuristic search of a finite space to find the most optimal solution" actually means
But isnt this proof that ML isn't residing in the PS5 and this method can be applied to just about any hardware without ML capabilities?
How does the PS4 version fair in the muscle deformation I am curious now
nope.
ML capabilities will run on anything. It's an algorithm like any other.
The training creates the algorithm, you still have to run it, which the PS5 is doing here.
It just takes horsepower to complete it, and particularly more challenging to get it complete in smaller frame times because the time to process a NN is fairly consistent regardless of it's inputs - so outside of changing the model, only having better hardware (which is a combination of compute and bandwidth) that will determine how much faster you can run it.
Actually I was indeed thinking that it is done in realtime in a network and then "bake" the information offline on the PS5.
So in the case of PS5 since it doesnt have any dedicated ML hardware, such tasks are taking extra performance, correct?
How did the PS4 version of morales differ?
No it doesnt, but ML can still be done on the machine. Its like RT being possible on the PS4 and Xone (crysis).
The same way XS may not have dedicated ML hardware. What it does have is reduced precision packed maths that can also be used for ML, but it can also be used for other tasks also.
I think people are getting hung up on this idea of dedicated.
The biggest difference will be perfomance, and if the reduced precision is a good fit for a particular task, including a particular ML inference.
D
Deleted member 86764
Guest
Is this fact or speculation?
Shouldn't ML be usable in anything that has int and fp processing power ? The difference being the amount of int and fp power brought to bare on the process ? So obviously a gpu with tons of cores will work better than a cpu with multiple cores.
I think the question isn't if the ps5 can do ML but if PS5 has any additional features for ML. I believe the suggestion is that MS has added more int 4 hardware (possible int 8 also ? ) above and beyond what is in the ps5 and that would give them an edge ? Because at the end of the day for any piece of hardware you will have to take from something to give performance some where else.
I'm not really in on the fore front of this conversation of ML in its self
The new muscles deformation tech is exclusive to the PS5 version.
The hardware in question (that we assume them to be) are mixed precision dot products which are useful for NN inferencing if you are mixing precision.
If you don’t have it you’d use a couple more cycles to obtain the same answer. How much an improvement it is will vary. But typically speaking, you’re not looking at a night a day difference like comparing regular compute silicon versus tensor cores, which are extremely dedicated to supporting the type of math NN use.
