vivster said:
Isn't the consistent drop in overall performance when increasing resolution a surefire sign that they did not go wide enough? For higher resolution you need to go wide because there is more stuff to do in parallel and they obviously did not go that route. Imagine where those cards were if not for the crazy clocks. Now the big question is how well does RDNA scale horizontally. If it does then Nvidia has a big problem on their hands because they are already quite wide and don't really have much potential to go much faster until they get on a new node. They will be forced to significantly remodel their architecture. Question into the room: How much do you think Samsung's node is at fault for the obscenely bad efficiency of Ampere? Will we be able to get double digit improvements on the Ampere refreshes on TSMC? If not, who is the most likely culprit for scratching at 500Watts? |
To be fair I'm not sure what else they could have done. They could make the top Big Navi die have more CU's but that would increase silicon budget in turn add cost, power and would need a greater cooling solution/s. And Navi 21 would end up a more expensive or on par 3080, but still with lower RT performance. About scaling, RDNA scales well with more CU's as we've seen in the lineup, thus far. Next generation of cards will be interesting with both Nvidia and AMD seemingly wanting to go with chiplet designs.
In any case AMD have met their targets with performance. And Nvidia, too with top tier performance across the board. Though they went with Samsung, which kinda threw a spanner in the works and allowed AMD to release a cards in the ballpark, which potentially can take some market share from them.
7nm will be faster as power efficiency gains will allow clocks to increase. I'm assuming Nvidia will release the big Super variants next year with 3080/3070 Supers. Akin to what they done with Turing's Super launches and potentially a 3080Ti. Not going to predict performance increase as its going to depend on Nvidia and if they are planning to have a stack ready to complete with AMD's new cards next year.