I think I was trying to get at that node shrinks no longer have double the transistor for the same price.
28nm was the last time that happened. 16nm is more expensive and 7nm should be more than 16nm.
Next Generation Hardware Speculation with a Technical Spin [2018]
- Thread starter Tkumpathenurple
- Start date
-
- Tags
- 5 bytes
- Status
28nm was the last time that happened. 16nm is more expensive and 7nm should be more than 16nm.
I don't expect any GPU manufacturer to sell high end parts directly nor I do I see a console manufacturer buying discrete parts from them either. I don't think the GPU manufacturer's markup is relevant to the console market as those markets are different (some overlap in high-end PC gaming and console, but not data center).
I think for AMD the benefit is they get a lot of help and funding for R&D that they in turn can use in their own products and how high end the console ends up is more on what the console manufacturer is willing to live with in terms of silicon cost.
Yes, that's absolutely true. Personally, I'm fine with next-gen consoles costing $499.
Tkumpathenurple
Veteran
I like that idea. Would it perhaps be better suited to the mid-gen model though? Release the base console with the most efficient use of silicon geared towards time to triangle, then include more exotic custom hardware in the mid-gen model.
They kind of did that with the Pro, by including custom checkerboarding hardware, and, I think, one or two other things that escape me right now.
It'd be a great focus for a mid-gen console too, since they wouldn't necessarily have to improve the rest of the system much in order to provide significant graphical improvement. That would then give them an easier time releasing the subsequent generation with substantially more power.
D
Deleted member 11852
Guest
Much of the ramping of cost below 20nm came through adoption of FinFET which had a host a challenges such a radical rethinking and design and chip layout. Looking at the process from a materials and physics perspective, 7nm should provide significant cost savings unless something unexpected happens. Thus far, nobody is reporting issues and if 7nm is in the next iPhone and/or iPad, those things are already being produced in serious volume.
And there's where you get cost savings, not re-designing chips on a 12/18 month cycle but making the exact same chip for a product that'll sell for many years, i.e. consoles.
No. You'll have a hard enough time getting devs to target your proprietary RT hardware if it's a platform exclusive HW feature. If it's only in 10% of the platform install-base, no-one will use it. RT hardware needs to be there from the beginning to have any shot of being utilised properly, plus it'll be a good marketing gimmick.
Why would you buy a new console if there is no generational gap new feature. I do think RT can be this new feature. As for the neural net thing, it could just be a cheap gimmick thing to inplement, like a tick box for marketting purposes, and it should not be very silicon intensive because you can even find them in current smartphone soc... They could shop around for IP from imgtec or arm althought they can get both neural net and RT from imgtec(i don’t know how scalable wizard is, and how it can fit in the AMD rendering pipeline) ...
Last edited:
The neural net could work well for denoising raytraced buffers. Once it has a meaningful justification beyond 'let's just throw one in there', it may be able to make a legitimate business case for its inclusion. Again, like RT, a piece of silicon that doesn't fit into developers models and won't be used on any other platform is only of value to a first party titles, and if it doesn't result in meaningful better results and sales, it's just a waste of money. $5 of extra silicon may not sound like much on a $500 console, but after $50 million consoles have been sold, that's $250 million you could be better off by not including it. It has to be worth more to the platform than that money spent elsewhere, like marketing or more GPU or a couple of big ticket exclusives.
I don't know maybe the extra 10+TFlops is a feature.
Neural nets in this case refers to machine learning where datasets are created offline and referenced on device. I don't think they'd be a great fit for AI agency. But that said, it's a new field and no-one really knows where they'll go. It'd be important to have the hardware freely usable if present. Same with ray-tracing. It'd be a shame if there was ray-tracing hardware but it was tied to visuals only and couldn't be used for AI, physics, audio, or whatever else the devs discover to do with it. Tying hardware to a purpose would be the very opposite of how GPUs have been extended to do all sorts of other work (although you do need to optimise your hardware for a task for it to be good at it and worth including over completely generic processing).
No.
Watch the PowerVR demonstrations to see what's possible. DXR on Volta is far from the only RT solution or option!
some form of neural net for animation could be useful.
much like how we use AI to 'fill in the blanks' for upscaling images etc. We might be able to have AI adjust the animation for a variety of scenes/effects that couldn't possibly be all modeled.
much like how we use AI to 'fill in the blanks' for upscaling images etc. We might be able to have AI adjust the animation for a variety of scenes/effects that couldn't possibly be all modeled.
Any links to that?
the old PowerVR stuff? Most of it was used for speed up development. But if you look up Power VR Ray Tracing hardware should find it.
From what I've read and probably some shreds of truth or fiction in here, the nvidia DXR hardware is pretty similar to what power VR did.
If there is ray tracing specific hardware, the gpu manufacturer would also support it with vulkan or other API by writing extensions. I suppose Nvidia needs MS in the mix because they don't allow vendor extension, but a console wouldn't have such restrictions in the way.
anexanhume
Veteran
If you ask developers what they want, it’s a faster horse. Cerny revealed with PS4 that developers weren’t interested in dedicated RT hardware or other eccentricies. They just wanted a balanced design with unified, fast RAM.
Similarly, once the current discussions about RT and DXR spilled onto Twitter, they made it (one of whom being @repi) clear they were envisioning next next gen when they were talking about it.
Similarly, once the current discussions about RT and DXR spilled onto Twitter, they made it (one of whom being @repi) clear they were envisioning next next gen when they were talking about it.
Last edited:
- Status
