I don't think the shown performance was from the topmost card!
There are several reasons:
- AMD never seems to start by showing off the best if there's more than one. (They didn't start by showing the Ryzen 9 5950X, but the 5900X.)
- They won't give Nvidia that much time to see the actual competition before the official presentation on the 28:th.
- They're unable to show the performance of the release cards because the clocks, power limits and voltages haven't been finally decided on yet.
Just like the RTX 3080 and RTX 3090 are almost realities when you visit a store to buy one off the shelf...
- AMD only provided somewhat detailed performance charts for 5900X and 5950X -- and the 5900X is the faster CPU for gaming. 5800X and 5600X were barely discussed.
- There's not much Nvidia can do at this point. Ampere GA102 is basically maxed out in the 3090, and Nvidia can't produce enough to meet demand as it stands.
- Final clocks might not be dialed in, but at this stage AMD is going to be within a few percent. Also, AMD isn't going to screw things up with last-minute clock adjustments like it did with RX 5600 XT.
As I note above, for AMD to get an 80% boost in performance over the 5700 XT is extremely impressive. Unless it has more than 5120 shader cores in Big Navi, that's far better scaling than anything Nvidia got from Ampere. Of course, Ampere's architectural changes are pretty extensive, so it's not that easy to compare. Let me give a few other data points, though.
RTX 2070 Super (2560 cores, 256-bit GDDR6) vs. GTX 1060 6GB (1280 cores, 128-bit GDDR5): 2070 Super is 125-150% faster. Double the cores, higher clocks, over double the memory bandwidth.
Titan RTX (4608 cores, 384-bit GDDR6) vs GTX 1070 Ti (2432 cores, 256-bit GDDR5): Titan RTX is 90-115% faster. Note quite double the cores, but triple the bandwidth.
RTX 2080 Ti (4352 cores, 352-bit GDDR6) vs RTX 2060 Super (2176 cores, 256-bit GDDR6): 2080 Ti is 55-65% faster. Double the cores, but bandwidth is only 38% higher.
RTX 3080 (8704 cores, 320-bit GDDR6X) vs RTX 2080 Ti (4352 cores, 352-bit GDDR6): 31% faster at 4K. Double the cores, roughly the same INT32 performance, and 52% more bandwidth.
Basically, if you want to double performance (or close to it), you have to double
everything -- not just the core counts. We don't know for sure what AMD is doing with the RDNA 2 architecture, and it's possible AMD will have separate INT and FP datapaths to boost performance (similar to Turing vs. Pascal). Still, assuming higher clocks, it looks like AMD is basically getting scaling similar to 2080 Ti vs. 2060 Super on the cores and RAM, and then the higher GPU clocks add maybe another 20%. That would be consistent with a 5120 shader core implementation. I think anyone who really believes AMD showed RX 6800 XT numbers will be in for a rude awakening at the end of the month.
To be clear:
Matching RTX 3080 is a huge deal for AMD. Last generation, at launch the RTX 2080 was 30-40% faster than AMD's best GPU of the time (Vega 64). And it used less power. When AMD pushed out the Radeon VII, the 2080 was
still 5-10% faster on average. With the RX 5700 XT, AMD improved performance per watt but the 2080 was still 10-15% faster. And by that time, the 2080 Super added another 10% to the Nvidia lead. So AMD had to compete on price.
This time, AMD appears to be able to come close to the RTX 3080 -- win some benchmarks, lose some others, probably using less power. That's an excellent improvement over the previous AMD GPUs. It would put AMD in the Zen 2 vs. Coffee Lake Refresh category as opposed to Zen 1 vs. Coffee Lake.