• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

Microsoft's DirectML is the next-generation game-changer that nobody's talking about

Claus Grimhildyr

Vincit qui se vincit
This isn't a game changer. It is a nice feature though.

Why the blatant overselling of this? Reeks of platform wars/fanboyism.
 

Klayzer

Member
Looks like it's just being reported, I don't see anyone calling this the holy grail.



He's replying to someone trying to shoot this down, it's just as bad as hyping it up. Worthy of being reported at least, is it not?
Yeah, "gamechanger" and "revolutionize" are not hyping at all. It's not like we haven't heard or read, the same buzzwords in almost every Microsoft article.

Ohh, and tlw is an Xbox hater and mega critic, but fans/antifans should wait a bit before definitive opinions. I myself will remain skeptical, until something concrete comes about.
 

Antwix

Member
I believe nvidia uses its dlss (or something similar) as its AI upscaling in the Shield correct? I've seen it first hand on the Shield and it is quite a cool feature.

But more importantly, I also just want to mention that the phrase "game changer" triggers me.
 

Claus Grimhildyr

Vincit qui se vincit
It's a big feature for a next gen console if supported.

Except it isn't a big feature. It is a minor one that can be nice, but it isn't necessary. When I think of "game changers" I think of features that are mandatory as they greatly change the landscape. This isn't one of them.
 

ZywyPL

Gold Member
Why would they need some secret sauce if they have 12TF?

Raytracing. You can have a 15-20TF GPU that will do 4K at 120FPS with ease, but then you apply RT and the framerate gets slaughtered. So staying at 1080p technically would allow to have a smooth RT performance, while on top of that leaving a lot of those 12TF for even greater details, instead of using them to just push native 4K like the X1X does, while at the end of the day still having that super detailed, crisp picture quality.
 

darkinstinct

...lacks reading comprehension.
Raytracing. You can have a 15-20TF GPU that will do 4K at 120FPS with ease, but then you apply RT and the framerate gets slaughtered. So staying at 1080p technically would allow to have a smooth RT performance, while on top of that leaving a lot of those 12TF for even greater details, instead of using them to just push native 4K like the X1X does, while at the end of the day still having that super detailed, crisp picture quality.
Yeah, machine learning upscaling makes a lot of sense when you want to have raytracing. Otherwise you would have to choose one or the other, 4K gaming without raytracing or 1080p with. Which means devs would have to optimize two different versions of the game just for one console.
 

psorcerer

Banned
Raytracing. You can have a 15-20TF GPU that will do 4K at 120FPS with ease, but then you apply RT and the framerate gets slaughtered. So staying at 1080p technically would allow to have a smooth RT performance, while on top of that leaving a lot of those 12TF for even greater details, instead of using them to just push native 4K like the X1X does, while at the end of the day still having that super detailed, crisp picture quality.

Unfortunately it doesn't work like that.
RT kills memory banwidth.
You can probably get away with cached pixels. But if your image reconstruction can access say 4 pix quad at a time it can do only as much...
But overall it can fly, if caches are overlapping. I suspect for any CNN the data duplication will be huge though...
 

ZywyPL

Gold Member
Unfortunately it doesn't work like that.

As DLSS is showing over and over again, it is. Also, bare in mind the next-gen consoles will load a lot of the data straight from the SSDs, leaving much of the actual RAM for whatever is needed. And the RAM configuration and specs are yet to be revealed.
 

psorcerer

Banned
As DLSS is showing over and over again, it is. Also, bare in mind the next-gen consoles will load a lot of the data straight from the SSDs, leaving much of the actual RAM for whatever is needed. And the RAM configuration and specs are yet to be revealed.

I'm not sure how DLSS is relevant here.
RT kills bandwidth, not amount of RAM.
 
I'm not sure how DLSS is relevant here.
RT kills bandwidth, not amount of RAM.
Having a single pipeline would allow less bandwidth, than having individual bandwidth allocations like on PC. This is why consoles should go along the route of having decided vram, ram, and cpu bandwidth, separate. Devs could do so much more.

Separate those, and RT will have less of a performance hit.
 

ZywyPL

Gold Member
I'm not sure how DLSS is relevant here.
RT kills bandwidth, not amount of RAM.

DLSS is basically the same AI-based image upscaling technique as DirectML. Secondly, the lower the resolution the lower bandwidth is required, as RT calculations are based on rays per each pixel, so for instance given a typical scenario with 3 rays/pixel the GPU's RT engine has to calculate 6.2MLN rays for each frame in 1080p, versus a massive 24,9MLN rays for native 4K, so yeah, that's A LOT of data.

But now combine the two technologies (like NV already does) - you lower the resolution to 1080p that isn't as much bandwidth (and computing) hungry as 4K in terms of RT calculations, and then you upscale the final image with DLSS, DirectML, CBR, or whatever the upscaling method you want.

As much as the very first incarnations of DLSS had very mixed results, to say the least, it's getting better and better basically each month, to a point today where it's not even hard/impossible to distinguish it from native 4K, but it even gives better results than the native image. And that Forza example with DirectML only further proves this is a worthy path, it's basically a win-win scenario where we get much better framerates and better picture quality at the same time.
 

darkinstinct

...lacks reading comprehension.
Having a single pipeline would allow less bandwidth, than having individual bandwidth allocations like on PC. This is why consoles should go along the route of having decided vram, ram, and cpu bandwidth, separate. Devs could do so much more.

Separate those, and RT will have less of a performance hit.
You can't do that until an engine is only using raytracing and not the hybrid approach anymore. That's something that will eventually be here - in 20 years st best. Until then, it is what it is.
 
You can't do that until an engine is only using raytracing and not the hybrid approach anymore. That's something that will eventually be here - in 20 years st best. Until then, it is what it is.
And until then... Would a limited single pipeline be beneficial, vs other proven methods? Isn't there bound to be a bottleneck at some point?
 

psorcerer

Banned
DLSS is basically the same AI-based image upscaling technique as DirectML. Secondly, the lower the resolution the lower bandwidth is required, as RT calculations are based on rays per each pixel, so for instance given a typical scenario with 3 rays/pixel the GPU's RT engine has to calculate 6.2MLN rays for each frame in 1080p, versus a massive 24,9MLN rays for native 4K, so yeah, that's A LOT of data.

But now combine the two technologies (like NV already does) - you lower the resolution to 1080p that isn't as much bandwidth (and computing) hungry as 4K in terms of RT calculations, and then you upscale the final image with DLSS, DirectML, CBR, or whatever the upscaling method you want.

As much as the very first incarnations of DLSS had very mixed results, to say the least, it's getting better and better basically each month, to a point today where it's not even hard/impossible to distinguish it from native 4K, but it even gives better results than the native image. And that Forza example with DirectML only further proves this is a worthy path, it's basically a win-win scenario where we get much better framerates and better picture quality at the same time.

I'm not sure it's relevant, again.
I have understood what you meant from the first post, no need to reiterate.
But you still haven't dived into my response.
Please think about how you will get the pixels that are required for image reconstruction.
 

ZywyPL

Gold Member
Please think about how you will get the pixels that are required for image reconstruction.

Why would I? Nor am I qualified for that, nor anyone is paying me to do so, while all the gaming-related companies out there do have their highly-qualified engineers to figure it out, and well, needless to say, they already did. It's been done over and over again, be it DLSS, CBR, Temporal Injection, Reconstruction, whatever, it's been done multiple times already, with better or worse results, it's already been proven to work, so I don't know what else do you want/need, what are you trying to prove? Just type in "DLSS" in YT and you'll have countless examples/evidences. It just works, and that's all I, the devs, and all other gamers care. Maybe you are some turbo nerd who needs to see the specific line of code or mathematical equations instead of the actual real-time results to finally believe, but of so, then it's just you and your own world then.
 

psorcerer

Banned
Why would I? Nor am I qualified for that, nor anyone is paying me to do so, while all the gaming-related companies out there do have their highly-qualified engineers to figure it out, and well, needless to say, they already did. It's been done over and over again, be it DLSS, CBR, Temporal Injection, Reconstruction, whatever, it's been done multiple times already, with better or worse results, it's already been proven to work, so I don't know what else do you want/need, what are you trying to prove? Just type in "DLSS" in YT and you'll have countless examples/evidences. It just works, and that's all I, the devs, and all other gamers care. Maybe you are some turbo nerd who needs to see the specific line of code or mathematical equations instead of the actual real-time results to finally believe, but of so, then it's just you and your own world then.

Again irrelevant.
I'm not arguing that DLSS doesn't work.
I'm telling you that RT random memory access exhausts bandwidth so fast that not much is left for DLSS. RT and DLSS will fight for resources and not help each other.
That's the only point.
 

RespawnX

Member
Again irrelevant.
I'm not arguing that DLSS doesn't work.
I'm telling you that RT random memory access exhausts bandwidth so fast that not much is left for DLSS. RT and DLSS will fight for resources and not help each other.
That's the only point.

That's exactly the point. RT needs a lot of bandwith and scales with the AI cores. DLSS doesn't need so much in comparison. That's why you combine both technices and split the resources, starting with a lower native resolution. At the end you get a similar picture quality to higher native resolution but you have more resources for RT and other calculation-intensive technologies. So you can use them to produce an overall better looking experience. That's why you want to use DLSS. Control and Minecraft RTX are perfect examples, highly recomend to take a look at the DF videos. Personally I still see a difference between native resolution and DLSS. But I also have to throw up every time I see checkerboard. DLSS ist far better than checkerboard renderding. At the end it's a question of how well the eye is educated for this kind of things. Only a few percent of people will notice the difference.
 

SenjutsuSage

Halo TV Series Promoter - Live from: Reach
DirectML seems to be Microsoft's answer to Nvidia's DLSS 2.0.

It will be interesting to see how Minecraft Ray Tracing for Xbox Series X compares to Minecraft RTX. Ultimately the high CU count was the right choice.

Yep, depending to what degree the Xbox Series X can take advantage of this and accelerate these operations, this is absolutely a game changer for next gen console gaming. All you need do is look at DLSS on Nvidia cards to see what's possible. Now obviously the Xbox Series X isn't packing nearly as much raw AI acceleration power as an RTX GPU, because there's no tensor cores. That said, we know that Xbox Series X fully supports DirectML and offers up 24 TFLOPS of FP16 for DirectML, as well as 97 TOPS of 4-bit integer performance. How far can that take Series X on next generation games? We shall see.
 

Deto

Banned
Checkboading used FP16.
Sony is the pioneer in these types of techniques with the PS4 PRO using checkboarding.

All companies should already have similar techniques, nvidia, sony, MS, AMD, Nintendo...

And obviously those who first used advanced techniques must not have stood by looking at the progress of others.

I find this hype in MS software bizarre, DX12 only exists because AMD launched the mantle and forced MS to update the DX.

While Sony and Nintendo are marketing games, consoles and accessories, MS is marketing the graphical API update.
 
Last edited:

dEvAnGeL

Member
The same way the cloud was the next game changer starting with crackdown 3? MS has always been full of shit when talking future tech.
 

Bolivar687

Banned
 

dxdt

Member
Checkboading used FP16.
Sony is the pioneer in these types of techniques with the PS4 PRO using checkboarding.

All companies should already have similar techniques, nvidia, sony, MS, AMD, Nintendo...

And obviously those who first used advanced techniques must not have stood by looking at the progress of others.

I find this hype in MS software bizarre, DX12 only exists because AMD launched the mantle and forced MS to update the DX.

While Sony and Nintendo are marketing games, consoles and accessories, MS is marketing the graphical API update.

Probably because MS added hardware support for 8-bit and 4-bit integer into the XSX SoC for Machine Learning. It's already being used to do automatic real HDR on games that were never designed for HDR. Second, DX12 Ultimate is now fully converged with PC, so the APIs is critical for Xbox and PC moving forward. They now share the exact same APIs and feature set.
 

Panajev2001a

GAF's Pleasant Genius
Definitely incorrect. 8.4 teraflops never came, hence all the checkerboarding we got. If this spin came from ms it'd get tore down every second lol. I find your obsession with me quite amusing tho.

At what rate does PS4 Pro process half floats? What would be PS4 Pro’s throughout if all operations were half floats? That is what Cerny said. Nobody, specially not Cerny, said PS4 Pro would not need checkerboard rendering especially in the same breadth as building HW acceleration for this.

Again, can you explain what the spin is supposed to be, where the incorrect statement lies and what was the lie told or what?
 

SleepDoctor

Banned
At what rate does PS4 Pro process half floats? What would be PS4 Pro’s throughout if all operations were half floats? That is what Cerny said. Nobody, specially not Cerny, said PS4 Pro would not need checkerboard rendering especially in the same breadth as building HW acceleration for this.

Again, can you explain what the spin is supposed to be, where the incorrect statement lies and what was the lie told or what?


Needed some time to run to Google lol. There's no need. Cerny also said all we need is 8 tf for 4k gaming. Our Ps4 pro "radically improves" performance to 8.4 tf, yet no true 4k gaming. 😁
 

Panajev2001a

GAF's Pleasant Genius
Needed some time to run to Google lol. There's no need. Cerny also said all we need is 8 tf for 4k gaming. Our Ps4 pro "radically improves" performance to 8.4 tf, yet no true 4k gaming. 😁

Ah, so you are now taking different unrelated statements and patching them together, context be darned, when the question was about explaining the reasoning you were apparently certain of?

Statement one was: to easily have current generation games quality at 4K you need 8+ TFLOPS GCN.
Statement two was about the rate FP16 values are processed at.
 
Last edited:

Bernkastel

Ask me about my fanboy energy!
The same way the cloud was the next game changer starting with crackdown 3? MS has always been full of shit when talking future tech.
You mean the "Power of SSD" or its predecessor "Power of Cell". You mean the PS2's "75 million polygons per second " lie. You mean that Killzone 2 demo that never looked like the actual game or Uncharted 4 and the promise of 60 fps(they even had trailer at 1080p 60 fps)
And who can forget the PS4 Pro reveal.
 

metaverse

Gold Member
You mean the "Power of SSD" or its predecessor "Power of Cell". You mean the PS2's "75 million polygons per second " lie. You mean that Killzone 2 demo that never looked like the actual game or Uncharted 4 and the promise of 60 fps(they even had trailer at 1080p 60 fps)
And who can forget the PS4 Pro reveal.

Agreed 100%. Microsoft delivered Crackdown 3 at 4K@60FPS, Sony has yes to do anything remotely comparable.
 
Top Bottom