DirectML is great but who’s going to create the actual models? Certainly not game devs. They have neither the skill set nor interest to doing so.
Speculation: GPU Performance Comparisons of 2020 *Spawn*
- Thread starter eastmen
- Start date
- Status
DirectML is great but who’s going to create the actual models? Certainly not game devs. They have neither the skill set nor interest to doing so.
DegustatoR
Legend
Well, I can see something along the lines of AI upscaling+AA running on DML being added into Unreal Engine and Unity at least. These should be used widely.
Its the models on one side and the hardware support on the other. DLSS 1.9 proves AI upscaling can be done through general shader cores but DLSS 2.0 shows that it can be done better on dedicated tensor like cores. Maybe someone else will show otherwise but as it stands it looks like NV are far better placed on the hardware front to handle heavy AI upscaling.
Microsoft and Khronos group would be the best positioned and most to gain from it. MS is best positioned because of their support for AI modelling and training and more importantly, the number of actual resources that work in the AI field. Google could support Khronos on this front as they also have a very experienced labour force in these types of fields.
You provide a model that is only allowed to work for Direct X/Vulkan and since the developers already support your API, they would incorporate a rendering path for it, or just make it a key component for rendering period.
Bondrewd
Veteran
MS and their army and then some of ML researchers.
Great!It's AMD. I *always* expect less.
And they've already pulled off some StyleGAN-ish stuff on Stadia didn't they?
The problem is around the difficulty of training these models. DML can be used for all sorts of game related things, its not just for resolution scaling so having it available doesn't mean its necessarily what devs will focus on using it for.
Consider that its taken NV 2 years to get something decent to the market and they were likely working on it long before DLSS 1.0 actually launched to the public. And its hard to imagine a company better suited to this given their direct access to world leading DL hardware, vast experience with game engines, and self contained hardware and software ecosystem.
Bondrewd
Veteran
mic-ro-soft.
Hope you're right and this number is >0. Putting some ideas up for use by interested parties did not work that well for AMD in the last years, so they had to go back to create ready-to-use packages you can download from GPU Open - and since they've done that, uptake has been increasing, methinks.
Let's see: „world leading DL hardware, vast experience with game engines, and self contained hardware and software ecosystem.“ Thats a no, no, no and yes in my books.
Bondrewd
Veteran
Stuff like project brainwave, yes, yes and kinda.
It's a hyperscaler bro, they're the 9000 pound gorillas of computing far beyond your understanding.
Sure they are a hyperscaler with Azure and more. But how much of that silicon did they develop themselves?
Jawed
Legend
I have no idea whether it's any good, but AI on FPGA via ONNX appears to be a thing:
https://docs.microsoft.com/en-us/azure/machine-learning/how-to-deploy-fpga-web-service
https://docs.microsoft.com/en-us/azure/machine-learning/how-to-deploy-fpga-web-service
It's not like MS is coming to the scene out of nowhere either, here's one of their talks at SIGGRAPH 2018 with DirectML "superresolution" upscaling based on model from NVIDIA, and I think it's pretty safe to assume they've worked on their own models before and since then, too.
https://on-demand.gputechconf.com/s...-gpu-inferencing-directml-and-directx-12.html
Frenetic Pony
Veteran
Uh, sure, gamedevs can do that no problem. Unity already has an ML agents plugin. Machine learning isn't some secret magic sauce only the highest level devs keep in ancient covens or some shit, hell most of its done on Python and a lot of it by people who just kinda fell into it at the right time.
What does developing silicon have to do with it?
You sure about that ?
MS has azure that already does machine learning and they have been doing it for years. MS has been in the game business since before Nvidia. Nvidia was founded in 93 just one year later ms started developing dx api.. However MIcrosoft Games Group released the first flight simulator in 82. So they have at least 38 years of gaming experience from designing games to the api they run on to actual hardware. We both agree with the last one so it doesn't matter for me to say much with that.
Then you have AMD who is no stranger to developing drivers , graphics api , computer hardware and game software . They also have a machine learning knowledge.
Then you have Sony. They have been in the game industry since the 90s.
All of them have a vested interest in getting ML running on RDNA 2 hardware as best as they can.
So yea its not just AMD in this its a lot of companies that would want this. IF MS is able to make it work they can allow it to work for AMD cards on pc and then MS will have an advantage over sony if they can't get something like DLSS running well. So I am sure MS would be happy to invest in it. Same reason why Sony would invest in it.
AMD will just reap the benefits
Most ML development in python is just people plugging in their data into models prepared by somebody else. This is about coming up with the model in the first place. That’s really the wheelhouse of “highest level” devs.
If any company is going to be able to replicate Nvidias results at the same or better quality in the next few years its most likely to be Microsoft. Although I still think they're at a disadvantage both technically and in terms of timescales (although collaboration with AMD could resolve the first of those disadvantages).
That does beg the question of what Sony will do though. They don't seem particularly well placed to compete in this area. I wonder if Nvidia would be willing to license their models to them? That would certainly have interesting implications for any PS5 games ported to the PC.
That does beg the question of what Sony will do though. They don't seem particularly well placed to compete in this area. I wonder if Nvidia would be willing to license their models to them? That would certainly have interesting implications for any PS5 games ported to the PC.
- Status
