Re. SLIDE AI / DL, I read that article with great interest at Tom's HW a couple of days ago. But it's really early days and no apparent commercial adoption (yet). Besides, there maybe room for both matrix multiplications (GPU) and hashing (SLIDE / CPU), depending on the specific data AI / DL tasks.
Be that as it may, I put a NVLink 2080 Ti 8k Superposition run in the spoiler...of course NVLink is not the same as single card etc etc etc and in any case, this was just a test run. The point is about potential efficiency
... dual 2080 Ti = 8704 CUDA cores and at 2.145 GHz, those sucked back a combined 760w (peak but fairly sustained) just for the two GPUs. If the 7nm 8192 CUDA GA100 can do this at 300w (or say even 360w with 120% PL), that would be an incredible step forward.
Finally, I don't think it is coincidence that this more detailed NVidia Ampere GA100 leak came out just when AMD had its Financial Analyst Day and presented RDNA 2
Yet with some serious impacts of covid-19 on the tech supply chains, who knows when we we actually get either AMD's RDNA 2 or NVidia's Titan or Ti versions of Ampere into our waiting (begging ?) hands.