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

yeah internet was more 1337 back in the days, i am not bbs old i jumped online quite late in the 90s cause my parents weren't wealthy enough to afford pc. But i remember the joy of configuring non plug and play modems/TCP stack and trying to get online on Win9x.
I am 1200/75 baud acoustic modem (think Wargames) Compunet service with a Commodore 64 and a black box to negate phone charges old.

I have to go outside now and shout at some clouds.
 
My online efforts at that time mostly consisted of being on AOL chat striking up conversations with young ladies...wonders if catfishing was a thing in the mid-90s.

I was behind the times. Had an interest in coding so took a BASIC class during the summer at the local community college and left after one day. I was like 14 and everyone else was like 8. When I looked around from my desk it felt like I had been held back 6 times and hadn't made it out of second grade. LOL
 
Only because they allowed people on the internet to teach it. Hmmm, what does that say about people on the internet? :p

Regards,
SB

That they are smart and figured out how shallow the AI was in a very short time, and then exposed that.
 
My first BBS exploring was done on a 300 baud modem on a Osborne 1. I still have it in storage. Tried AOL a few times later on my custom built 286 using a 2400 baud modem, but local BBS were still my favorite. It wasn't until my 3rd year of college that I really started going online with something called the Internet. By then I was using 14.4 & quickly upgraded every time a new speed was available. My last dial-up was 128k ISDN before going to cable for good. Man those were exciting times. But now I feel old. :(

Tommy McClain
 
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.
Does PS5 GPU support double-rate FP16 that resolves into FP32 results?

Vega RPM FP16 resolves into FP16 results.

RX 6800 series' double-rate FP16 can resolve into FP32 results.
 
"ML stuff" is essentially a bunch of compute kernels doing maths for you, and your typical latest-gen GPU can already run them great with stuff like packed FP16 and Int8 support. Assuming there is no CU capability divergence given both being RDNA 2, the only possible difference is actually backing some of these kernels with accelerators (e.g. "tensor cores" for running convolution and friends).
ML instructions are optional features for NAVIs.

Don't assume Vega RPM is the same as the optional NAVI mix data format capable RPM.
 

Attachments

  • NAVI RPM and Dot HW support Feature test.png
    NAVI RPM and Dot HW support Feature test.png
    424.6 KB · Views: 13
I'm liking Sony's ideas on use of AI tech. These types of solutions could eventually reduce or remove the need of mocap animations. Would be great to see more blended and natural reactions to changing environments.
 
I'm liking Sony's ideas on use of AI tech. These types of solutions could eventually reduce or remove the need of mocap animations. Would be great to see more blended and natural reactions to changing environments.
this is largely where I expected to see a lot of ML happen. Animation and being able to connect those points of say a body properly with the environment in a fluid way. This is much more important than DLSS imo at least in the console scene. We can get away with other methods of reconstruction because that's not what consoles are about really - but polishing the game details themselves is where I want to see ML attention focused.
 
That is example of ml usage, this topic is not only about confirmation for mix data capabilities
ML can run on any DX11 GPU, the debate is about efficiency i.e. which hardware is more suitable to machine learning.

Any DX11 GPU can execute lesser data types for ML but it's without double rate or quad-rate or octa-rate processing.

NAVI 10 baseline IP has double rate INT16 and FP16. The double rate INT16 feature can process INT4 and INT 8 at a double rate processing.

Microsoft is claiming Series X offers 49 TOPS for 8-bit integer operations and 97 TOPS for 4-bit integer operations.

NVIDIA's RTX Tensor cores are the ideal hardware ML.
 
ML can run on any DX11 GPU, the debate is about efficiency i.e. which hardware is more suitable to machine learning.

Any DX11 GPU can execute lesser data types for ML but it's without double rate or quad-rate or octa-rate processing.

NAVI 10 baseline IP has double rate INT16 and FP16. The double rate INT16 feature can process INT4 and INT 8 at a double rate processing.

Microsoft is claiming Series X offers 49 TOPS for 8-bit integer operations and 97 TOPS for 4-bit integer operations.

NVIDIA's RTX Tensor cores are the ideal hardware ML.
yeah, I think almost everybody in this topic know this facts ;)
 
I'm still waiting for a game to actually show that this topic is even worth discussing ... But in the meantime, we should probably agree that there are three categories of neural network ML use on consoles:

Neural Network Models
1. Use the end-product of NNML in a game, e.g. an animation, even when several are blended, that was generated from an ML model, but is static data in the game.
2. Use the ML model in a game, e.g. an animation is generated applying a previously (cloud) trained ML model. An animation can be controlled from this model, but it will still always generally behave the same way in the same situation, but it can handle just about any situation old or new, from the parameters and rules calculated from the model that was derived from the training model.
3. Use actual ML training model in a game, e.g. there is an adaptive neural network in the game code that learns from input from the player, all players, in any shape or form

From there, we can discuss whether specific hardware and/or API features help any of these categories and how, and how they fit into games.

Additionally, there are other forms of machine learning that do not use Neural Networks, that we should probably also discuss.
 
FIFA 22 uses some type of trained machine learning animation system that they're calling "Hyper-motion". It's available on Series X/S and PS5.

https://www.ea.com/games/fifa/fifa-22

MACHINE LEARNING
A cutting-edge proprietary machine learning algorithm learns from over 8.7 million frames of advanced match capture, then writes new animations in real time to create organic football realism across a variety of interactions on the pitch.
 
Last edited:
I am still wondering what ML features does Xbox have. It was marketed at one point but now it is kinda forgotten. Considering that we hear more and more about ML and Playstation 5 is also there, I wonder if ML support in Xbox is a nothing burger.
 
Back
Top