NVIDIA Custom graphics cards

[steve10765]

I am not sure if it is hard to make graphics cards or not, I am assuming they are, seeing as you can't order custom graphics cards, and it may be pointless seeing as games might have to be coded for what I have in mind...

Nvidia should make custom cards, I was thinking a graphics card with 15,000 CUDA CORES clocked at around 850MHZ with respectable shaders etc... This would obviously cost a sh!t ton of money, priced at about 9,999 ( my thoughts were 3000cuda cores in 690 and thats 1k, so to have 15k cores it would cost 5k, but because this is probably hard to make it cost more and they want more profit)

Could a game even use this? I think you might need to understand how GPU coding and optimizing games works, I doubt it is like this but I was thinking it goes by the amount of cuda cores, so say a game detects 13000 cuda cores it will transport the load between those cores.

explain your thoughts...
I was recently reading this thread and it made me think of this

 

[Oreo]

There are so many things preventing them from doing this.

Right to start off straight away assuming you could even do 15,000 cuda cores on one chip, the bandwidth and power required are likely to outstrip PCI-E 3.0 and PCI-E power connectors.

Secondly, this thing is going to be ****ing toasty warm. You'd need a cooler the size of a 2 Liter bottle of soda to keep it cool.

Thirdly, it's going to be monstrously big. No motherboard is going to fit it.

Forth, there are so many technical issues in designing a chip with this many 'cuda cores' and high clock rate. The rate at which the silicon wafers would come out alright would be TINY. You would be throwing away 10,000 wafers for every one that comes out okay.

You can't just simply adding cores and increasing the clock rate. All the other parts have to be upgraded as well. It's like putting a huge engine in a small car, it would destroy it, you have to upgrade the gearbox, drive shaft, wheels, tyres etc. Same with a graphics card. You need to put some huge ass bus on it, some very quick memory, insane memory controllers and that kind of thing.

Could it be done ? probably. But for $10,000 ? ahaha. You got to be kidding. Increase that by a factor of 20 and we might be getting somewhere more realistic if this was going to be a profitable proposition for nvidia, and that is assuming there was even a market for them.

 

[carnageX]

Could it be done ? probably. But for $10,000 ? ahaha. You got to be kidding. Increase that by a factor of 20 and we might be getting somewhere more realistic if this was going to be a profitable proposition for nvidia, and that is assuming there was even a market for them.

This. Research and development for such a thing would require this to be very high, because it would be geared towards such a limited market.

 

[PP Mguire]

It's not a case of any of that, because we will eventually get to that point. It's a case of if they could right now within the thermal and power specs while maintaining a respectable chip size they would. They just simply can't make a 15k core chip right now due to limiting factors in manufacturing right now.

Edit: A good analogy is if it was cost effective for everybody to have super computers then they would.

 

[Oreo]

Well everything I said was based around current manufacturing proccesses and technology, of course.

When we are knocking on the door of 4nm i'm sure it won't be quite so hard to do all the above.

 

[steve10765]

thanks for the reply's people, and that analogy by kmanx helped alot

 

[Apokalipse]

Forth, there are so many technical issues in designing a chip with this many 'cuda cores' and high clock rate. The rate at which the silicon wafers would come out alright would be TINY. You would be throwing away 10,000 wafers for every one that comes out okay.

It would have less than 1.7% yields

Actually, with 15,000 GPU cores, how many GPU's would even fit on a wafer?
GK104 has a die size of 320mm² with 1536 shaders, so multiply GK104's die size by about 10, and you get 3200mm² per GPU

 

[Oreo]

It would have less than 1.7% yields

Actually, with 15,000 GPU cores, how many GPU's would even fit on a wafer?
GK104 has a die size of 320mm² with 1536 shaders, so multiply GK104's die size by about 10, and you get 3200mm² per GPU

Ha.

How did you get the 1.7% yield figure ?

 

[C0RR0SIVE]

That would be one big ass die... o.0

 

[PP Mguire]

That would be one big ass die... o.0

That's what she said?