RED.Misfit:

Lisa is hopeful that in the 2nd part of 2022, shortage will ease down. For myself, i see light in China banning cryptomining and "Ethereum 2.0" to lessen the volume of GPU used for cryptocurrency. When retailer (or whatever intermediary is trying to rip our pockets!) will have full high price GPU they can't sell, they will lower price.

The best when the old VGAs will flood the market. As soon as the new generations arrive there will be hundreds of VGAs for sale. I can imagine how many nvidia cards are now working in mining rigs. Of course if the new generations will be low hash rate ones then many will keep their Ampere cards until better comes up.

Devid:

The best when the old VGAs will flood the market. As soon as the new generations arrive there will be hundreds of VGAs for sale. I can imagine how many nvidia cards are now working in mining rigs. Of course if the new generations will be low hash rate ones then many will keep their Ampere cards until better comes up.

That is the reason why I can sell my 5700 now and get a 6700xt or even 3060ti because of the hash rate This 1 is going on ebay later today

tunejunky:

a few educated guesses.... 1) 3D- V Cache shows up

X2

tunejunky:

2) definitely wider memory bus, certainly if there's 3D- V Cache

Don't you mean the opposite? If they add L3 cache, they won't need to increase memory bus width.

tunejunky:

and remember this is a smaller process AND no-one is more adept at mcm than AMD which leads to ... 3) a separate multicore tensor / RT chiplet married to the controller remember the only limits to Infinity Fabric implementation on a pcb are imagination and socket size

I agree they will go full on chiplets for RDNA3. But I think they won't have tensors and RT units on a different chiplet, because that would cause a latency spike for the rendering pipeline. I bet they will have a complete chiplet, with shaders, cache, ROPs, TMUs, RT units, tensors, etc. Then they just add chiplets as much as they want, creating a product stack, from low to high end. Remember that RDNA2 is a tile based render arch. So it can have one chiplet render one tile, and another rendering the next and so on.

Horus-Anhur:

X2 Don't you mean the opposite? If they add L3 cache, they won't need to increase memory bus width.

cache is good and all, but you still need solid memory subsystem performance to back it up at higher resolutions, that is clear even on the current rx 6000 series products.

i never get how "next gen card is faster than this gen card" articles are news.