AMD Vega Discussion Thread

[the54thvoid]

Nope , Vega still supports the same features previous GCN architectures had it just has extra stuff and handles some things differently. Read carefully what I wrote.

Pascal for example has improved context switching aka is more asynchronous compute capable than Maxwell which isn't at all to more precise.

They are compatible but that works only 1-way unless you want performance degradation.

No slam dunk.

The idea is Nvidia's architectures have been segmented throughout the years in terms of features and capabilities.

I slammed so hard the back board smashed.

 

[BiggieShady]

Maxwell still uses static scheduling , that element alone makes it quite different from Fermi...

Yeah well Kepler, which I incidentally forgot to mention, brought back static scheduling back in a fight for efficiency and it worked well for gaming workloads ...
... which of course has nothing to do with structural progression of Cuda arch which I was referring to ...
... am I crazy (don't answer, I know I am) or nvidia in the year 2035. will have a GPU that covers the die as a fractal ...

 

[Vya Domus]

... am I crazy (don't answer, I know I am) or nvidia in the year 2035. will have a GPU that covers the die as a fractal ... View attachment 91612

I wonder how long will it be until these companies will be forced to use only complex optimization algorithms for designing these dies and make them look like that because that would be the only way they'll be able to squeeze more efficiency reliably.

I mean I know some elements are designed that way but I am convinced the bulk of the work is still done by hand.

 

[BiggieShady]

I wonder how long will it be until these companies will be forced to use only complex optimization algorithms for designing these dies and make them look like that because that would be the only way they'll be able to squeeze more efficiency reliably.

I mean I know some elements are designed that way but I am convinced the bulk of the work is still done by hand.

I was thinking hilbert curve has a nice covering for rectangular die when modular hierarchy is needed:

I went off topic a bit, but yeah they'll use their deep learning neural networks for bulk of the optimization and do important stuff by hand (maybe even steal my ingenious idea from this very post )

 

[Zubasa]

Apparently flashing a Vega 64 bios on my 56 got me a Vega 57.

Edit: According to the latest news on TPU, it is a bug