This can be enabled on most amd cards through custom r.id drivers.
The 6800xt has it enabled by default on normal drivers as part of its feature set.
(nv also has something similar called DMC but it needs to be on application side to be used.)
In my limited tests in past rad7 even with compression wasn't able to hit 1TB/s
it was something around 820GB/s, and without compression was some 620-710GB/s (from my memory), there was around 5% performance gain at most. *(even so in my tests fury, vega, and rad7, and nv side was always starved of memory, and cause was high latency on hbm)
7900xtx it was hitting around 1.8TB/s, and there was actual regression in performance (single digits), it was also causing higher than norm idle temps - so i just kept it disabled - there was little to no performance difference with 7900xtx when you downclock memory so card performs about the same with 650GB/s as with stock "960GB/s" that it can't reach anyway. Thus that specific gpu isn't as reliant on fast memory - and mostly hitting some other bottlenecks inside.
Thus question is - is it beneficial to your card, and you need to test that; from my exp its a mixed bag.
Most people think the card is starved from memory throughput, and think if it had faster memory it would be faster.
It doesn't really work out this way in most cases in my opinion (so you cannot look through prism of nv 'super' cards, as those cards typically have better wafer, more cores and so on allowing them to benefit from better memory).
*(this was same for all generations of 'super' cards. They sometime got faster memory, but always their Shader/TMU/ROP count has changed.)
github.com
