Support NeoGAF

[LivingD3AD]

Actual topic here: https://www.neogaf.com/threads/inso...r-character-deformation-in-real-time.1598777/

 

[Negotiator101]

We already know this.

 

[Ailike]

BLAST PROCESSING! EMOTION ENGINE! BUZZ WORDS!

 

[Abriael_GN]

Literally everything with a chip is "capable" of machine learning.

 

[reksveks]

Again Joe seem kinda stupid or has a tendency of hyperbole, we don't know how computationally expensive this process is.

 

[Abriael_GN]

Again Joe

Are we supposed to know who this is?

 

[v_iHuGi]

Josh is a technical Lead at Insomniac, he's on a very high position there.

Thank me later.

Ps5 is a :

 

[Abriael_GN]

Josh is a technical Lead at Insomniac, he's on a very high position there.

Great, but Joe is not Josh.

 

[reksveks]

Are we supposed to know who this is?

Not really, just some weird Sony 'fanboy' imo whose tweets are kinda cringey.

 

[Abriael_GN]

Not really, just some weird Sony 'fanboy' imo whose tweets are kinda cringey.

Thought so, thank you.

 

[Shmunter]

Rise of the machines

 

[Haggard]

imagine if people had this kind of talk about every other budget pre-built PC some discounter puts out into the open.....

 

[onesvenus]

All GPUs are capable of running ML, the question is if they have specific hardware to accelerate inference.
I'm sure PS5 has at least the basic acceleration found in RDNA2 GPUs

 

[Onironauta]

Any computer can run machine learning algorithms. The doubt is whether PS5 GPU supports hardware acceleration for it.

 

[archangelmorph]

I mean this is true through any lens...

But it's important to understand the distinction between training an ML model (which is where the bulk of the computation is) vs using it to make predictions (in this case being resolving human-like muscle geometry transformations given an arbitrary model mesh & skeletal rig.

The former would be done offline using likely large cloud-based neural networks.

The latter would be in this case performed in real-time.

 

[M1chl]

ML model can be run on anything, does not say if they have dedicated block of logic to accelerate that....well based on what we know RDNA 2 does not really have something akin to Tensor Cores. But it's nice, that devs are utilising this.

 

[jroc74]

All GPUs are capable of running ML, the question is if they have specific hardware to accelerate inference.
I'm sure PS5 has at least the basic acceleration found in RDNA2 GPUs

Any computer can run machine learning algorithms. The doubt is whether PS5 GPU supports hardware acceleration for it.

I mean this is true through any lens...

But it's important to understand the distinction between training an ML model (which is where the bulk of the computation is) vs using it to make predictions (in this case being resolving human-like muscle geometry transformations given an arbitrary model mesh & skeletal rig.

The former would be done offline using likely large cloud-based neural networks.

The latter would be in this case performed in real-time.

ML model can be run on anything, does not say if they have dedicated block of logic to accelerate that....well based on what we know RDNA 2 does not really have something akin to Tensor Cores. But it's nice, that devs are utilising this.

I could have sworn there were some (many?) that said the PS5 couldnt do it at all. After folks explained it was there in RDNA 2 gpu's. Because some are still riding on the PS5 is RDNA 1 train.

Yeah, it might not be on the same level as MS, who knows. But this wasnt what the argument was at first.

 

[M1chl]

I could have sworn there were some (many?) that said the PS5 couldnt do it at all. After folks explained it was there in RDNA 2 gpu's. Because some are still riding on the PS5 is RDNA 1 train.

Yeah, it might not be on the same level as MS, who knows. But this wasnt what the argument was at first.

It was speculation as far as I know, before we know how even RDNA 2 operates. But still would be nice to have something like Hot Chips for PS5.

 

[longdi]

I could have sworn there were some (many?) that said the PS5 couldnt do it at all. After folks explained it was there in RDNA 2 gpu's. Because some are still riding on the PS5 is RDNA 1 train.

Yeah, it might not be on the same level as MS, who knows. But this wasnt what the argument was at first.

all modern gpu can do ML, just less efficient if you lack dedicated ML hardware as is the case with ps5.

insomniac probably have excess capacity seeing MM is a last gen port

 

[Griffon]

As usual, tech illiterates making up dumb shit.

They use machine learning in their studio to set up the custom weights and deformation code on the character models. Once it's running in the console there's no learning happening.

 

[jroc74]

all modern gpu can do ML, just less efficient if you lack dedicated ML hardware as is the case with ps5.

insomniac probably have excess capacity seeing MM is a last gen port

This wasnt the initial argument, you know this.

Whats happening now is ppl are now diving into the specifics of how its done. Initial argument was it couldnt do it at all on any level.

 

[v_iHuGi]

Great, but Joe is not Josh.

It was Josh who confirmed, it's right their.

 

[Physiognomonics]

As usual, tech illiterates making up dumb shit.

They use machine learning in their studio to set up the custom weights and deformation code on the character models. Once it's running in the console there's no learning happening.

Who is tech illiterate?

Not sure how specific I can get with specs for now, but you are correct in the assumption that all final deformations are resulting via ML inference at runtime on the PS5 hardware. There are no blendshapes, skin decomp, or traditional tricks of the trade (nothing against them!)

 

[longdi]

This wasnt the initial argument, you know this.

Whats happening now is ppl are now diving into the specifics of how its done. Initial argument was it couldnt do it at all on any level.

what initial argument?
we know SX has special ML sauce while ps5 don't.
doesn't mean ps5 can't do ML.... it's just going to be less performant.

physics is physics.

 

[Md Ray]

Josh is a technical Lead at Insomniac, he's on a very high position there.

Thank me later.

Ps5 is a :

He was also involved in the VFX of The Amazing Spider-Man 2 movie!

 

[ScaryBrandon]

Console fanboys are so cringey since the vast majority of the time they try to talk about things they don't understand.

It's forgivable when you are a console kid but how do adults get to that point?

 

[Griffon]

Who is tech illiterate?

I confirm, tech illiterate.

ML Inference is not live ML. It's the algorithm resulting from ML, generated at the studio. The console does nothing but execute the resulting regular-ass code.

 

[Physiognomonics]

what initial argument?
we know SX has special ML sauce while ps5 don't.
doesn't mean ps5 can't do ML.... it's just going to be less performant.

physics is physics.

Just like 12 > 10 right? PS5 is going to be less performant in games?

 

[azertydu91]

Literally everything with a chip is "capable" of machine learning.

No I have a muffin with chocolate chip and honestly I have been disapointed on that front.

 

[longdi]

Just like 12 > 10 right? PS5 is going to be less performant in games?

yes probably once next gen engines take over

 

[Md Ray]

all modern gpu can do ML, just less efficient if you lack dedicated ML hardware as is the case with ps5.

insomniac probably have excess capacity seeing MM is a last gen port

what initial argument?
we know SX has special ML sauce while ps5 don't.
doesn't mean ps5 can't do ML.... it's just going to be less performant.

physics is physics.

It's "running as fast as the non-physically based version". Meaning the perf cost, in his words, is non-existent when running this tech using PS5's ML.

I know this news has you shook. But come on now.

 

[longdi]

It's "running as fast as the non-physically based version". Meaning the perf cost, in his words, is non-existent when running this technique using PS5's ML.

I know this news has you shook. But come on now.

as i said, spare capacity from being a last gen port

 

[v_iHuGi]

Console fanboys are so cringey since the vast majority of the time they try to talk about things they don't understand.

It's forgivable when you are a console kid but how do adults get to that point?

Most don't even know to write a simple line of code, let alone talk about advanced stuff like this.

I mean how many people understand this simple loop?

for (var i = 1; i<10;i++) {
console.log(i);
}

 

[Physiognomonics]

I confirm, tech illiterate.

ML Inference is not live ML. It's the algorithm resulting from ML, generated at the studio. The console does nothing but execute the resulting regular-ass code.

On Nvidia GPUs, DLSS is live ML?

 

[jroc74]

It's "running as fast as the non-physically based version". Meaning the perf cost, in his words, is non-existent when running this tech using PS5's ML.

I know this news has you shook. But come on now.

Its amazing seeing goalposts move. I knew this was coming, lmao.

 

[Schmick]

Just so we all can be clear on what Machine Learning Inference is

Machine Learning Inference

Machine learning inference is the process of running live data into a machine learning algorithm to calculate output such as a single numerical score.

hazelcast.com

 

[longdi]

Its amazing seeing goalposts move. I knew this was coming, lmao.

similarly, im stun at how insistence some are pushing the agenda that ps5 is of equals to SX. when even both the consoles marketing don't suggest parity

 

[Shmunter]

as i said, spare capacity from being a last gen port

Maybe. But retrofitted last gen code as opposed to tailored next gen code taking ML into account from the outset could actually be less efficient.

 

[Md Ray]

as i said, spare capacity from being a last gen port

Last-gen port or not, if PS5 lacked hardware for ML then there should have been a hit to perf.

You're in full-on damage control mode.

 

[azertydu91]

Most don't even know to write a simple line of code, let alone talk about advanced stuff like this.

I mean how many people understand this simple loop?

for (var i = 1; i<10;i++) {
console.log(i);
}

It shows in the console an increment of the varaiable i, meaning you will see all numbers from 1 to 10.

 

[longdi]

Maybe. But retrofitted last gen code as opposed to tailored next gen code taking ML into account from the outset could actually be less efficient.

nah this sounds like a fun side project and early research spillover. seeing its only applied on spidey

you can only do that with spare capacity. why would your boss allow youto break the established with rather 'pointless' addition

 

[Xyphie]

No one has ever argued that PS5 can't do machine learning, because you can do machine learning inference on any modern GPU or CPU. What people are arguing about is what data types (FP32, 2xFP16, 4xINT8, 8xINT4) can be packed concurrently.

Sony has no one to blame but themselves for this because they've been terrible at providing technical details on the actual feature set of their GPU (IMO, likely because it's less extensive than the Xbox GPU).

 

[Shmunter]

similarly, im stun at how insistence some are pushing the agenda that ps5 is of equals to SX. when even both the consoles marketing don't suggest parity

In all fairness do we have a practical example of on gpu ML running on XsX?

 

[v_iHuGi]

It shows in the console an increment of the varaiable i, meaning you will see all numbers from 1 to 10.

Oh damn you told the answer to the other 99%!

 

[azertydu91]

Oh damn you told the answer to the other 99%!

Hey I've been in an IT school while being high and fighting alcoholism and depression, I wanted to see if anything remained.

Edit: By the way what remained is liver and bladder problems ...That was not worth it.

 

[ozz ex machina]

Why the fuck is there a thread for this.

 

[Agent_4Seven]

Yeah but... will any 1st party games actually be using touchpad this gen for something other than huge map button yet again?

 

[FranXico]

I confirm, tech illiterate.

ML Inference is not live ML. It's the algorithm resulting from ML, generated at the studio. The console does nothing but execute the resulting regular-ass code.

And how do you think DLSS works?

 

[jroc74]

similarly, im stun at how insistence some are pushing the agenda that ps5 is of equals to SX. when even both the consoles marketing don't suggest parity

Still relying on marketing....huh...

No one has ever argued that PS5 can't do machine learning, because you can do machine learning inference on any modern GPU or CPU. What people are arguing about is what data types (FP32, 2xFP16, 4xINT8, 8xINT4) can be packed concurrently.

Sony has no one to blame but themselves for this because they've been terrible at providing technical details on the actual feature set of their GPU (IMO, likely because it's less extensive than the Xbox GPU).

No, there are ppl who said this.

What you posted is the answer ppl gave saying it does. Only the more logical, informed ppl agreed with the latter.

Again, goalpost are a moving.

Even if Sony didnt give any details at all, that means ppl should immediately think the worst about the console?

 

[turtlepowa]

Why the fuck is there a thread for this.

Because console wars, it doesn't matter that 90% don't understand what it's all about.