Mathematical approach to future GPU designs?

[ewilson248]

Note: This is a theoretical question not based on the current real world or any existing or soon to be released hardware. This is about applying existing facts on graphics specs in a different way.

 

Has anyone seen articles or calculations that provide VRAM capacity requirements or hardware bandwidth requirements based on a certain set of theoretical required output goals? Could this be reduced down pure math based on hard facts? The number of pixels, color depth, frames per second, Memory capacity and bandwidth, GPU bandwidth etc.  In plain speak: If your design goal was the following: 16 bit color, 8k resolution at 120 FPS, 200w TDP  what would be needed? What other factors should be considered?

 

Today's merger of ARM and Nvidia has me pondering a drastically different design approach to graphics cards in the future that do NOT involve massive size dies sizes and 350w TDP!

Perhaps scalable chiplets?  If we were to start with a blank sheet from our output goal what does the math say we need to accomplish that?  

 

 

[Zat0ichi]

Stop being intersting. This is the internet.

 

I get your point though. You set an almost console rigid standard so people get the most efficient product that does not have buckets of spare overhead. We're paying for that over head.

New ddr6x is expensive. 4k gaming is only ever going to need so much. Why pay for more than we're ever going to use. You need just a bit more horsepower than 3070 but not as much as the 3080.

 

A mix and match build you own GPU.

 

Nice idea.

 

A scanner for you system that will then customise any part of you system with the optimum component.

You could add a price filter then use sliders on various stats to balance your own product.

The technological ecosystem that would be in place for that to be possible would make the concept of discrete component gaming obsolete. Neural laces in Peter f hamilton?

 

Even though chip fab is just lithography the templates have the yield issues. That's the reality stumbling block. Custom printing chips rather than churning out the same one hoping to get a percentage that works.

 

 

 

 

[ewilson248]

I am speaking in terms of a chip manufacturing design process only.

 

If NVIDIA went to a modular chip design where smaller dies could be used in multiple sets, net yield would go up and the number of SKUs would go down?

Throwing larger dies and more power at the problem is not an answer long term. Then you use a building block approach to custom tailor the board design based on a price goal, a performance goal, or power goal.

 

Doesn't the PlayStation architecture or some other CPU / GPU already do this type of modular thing?  I remember this concept was already adopted by someone. Why not die shrink a Jetson GPU to 7nm and start from there? My guess is with ARM in the corporate IP portfolio now, massive multi chip designs will be the wave of the future for team green. 

 

[Stockholmes]

I think if it was that easy they would've maybe done something like that already. I don't think it's that easy. Mixing and matching a GPU together from many different elements that is scalable to what you want? You know that these companies spend *a lot* of money just researching & developing the (premade) chips we consumers buy? Then again I know jack**** about manufacturing GPUs so don't take my word for it. I just think it can't be that simple..

[Zat0ichi]

There are already cores within GPUs.

Stitching those together is the key mechanism you need.

So a core hub on die.

 

Set you usage parameters and it will print the prerequisite number of cores. Still changing the print template at the scales involved on a chip by chip basis is a few years off. Single chip printing?

 

The sticking it to a PCB is relatively simple.

 

7nm fab is crazy small and delicate