Ah you're saying next gen nVidia won't have any real compute benefit with it's own InfinityCache version but should see gamer speeds improve? They might put that cache on the supers - I don't know how difficult it would be to integrate that into the design. I imagine they'll want the Super out sooner because AMD cards have the RAM advantage at the time being and I think the Supers were slated to have double the VRAM, so maybe no cache on them.
I think an AMD compute card coming out next March would probably blow the 3080 and Radeon VII out of the water if it had HBM and 512 bus. Considering stock is so limited now looks like I'll be forced to go that option. I don't like only having 1 or 2 cards of a certain type - much easier to populate a rig with 6-8 of the same card.
The question is, what is cost benefit here, cache or a better gram? AMD used cache and a common gdddr6, nvidia used the top gddrx6 ram. I wonder what would have happened if amd used cache and a top gddrx6. If amd wanted to create a eth mining gpu they have the means, instead of cache add hbm or both.
I think they are adding HBM to their compute card that is coming out. They know they need to compete with nVidia in the gamer market because it will help sell their CPUs at the same time. nVidia probably has more sway commanding better pricing for GDDR6X than AMD did, but AMD has the history with TSMC and is sitting pretty with that 7nm node. With SAM the 6800 is best performance per $ in gaming at present, but meh for mining.
Datacenters already use EPYC for their clusters but with nVidia GPUs. With a good compute card they might be able to take some of that deep learning market share (and make miners happy).