@JarredWaltonGPU why do you think the 6800xt is so close to the 7800xt? And do you think AMD spent a lot of money designing these new cards, or just took this generation off?
I think AMD put a lot of effort into reworking everything for their GPU chiplet architecture. From a performance standpoint, it didn't pay off except for RX 7900 series, because those have more Compute Units (CUs). But chiplets are inherently less efficient and so it hurts there as well.
On a generational basis, 6800 XT had 72 CUs clocked at around 2.25 GHz (often a bit more, but that's spec) while 7800 XT has 60 CUs clocked at 2.43 GHz. Architecturally, the only massive change is the doubling of compute potential, but it seems that's not really leveraged by games. I'm not sure why that is, but I suspect it's partly that the doubling of compute isn't always useful, and partly just an inability to keep all the compute elements fed with data.
AI (Stable Diffusion) is the only thing where I've really noticed a big jump in RDNA 3 vs. RDNA 2. And note that the "AI Accelerators" in RDNA 3 are just sharing the compute shader stuff, which was also done with RDNA 2 but just wasn't talked about much. Anyway, I think there's more support for this in RDNA3, so that you can issue one "AI instruction" instead of several "shader instructions" to accomplish the same task.
Nvidia's Ada isn't a radical change from Ampere, architecturally (outside of the DLSS 3 / Frame Generation stuff, which is a whole other can of worms), but things were reworked to get much higher clocks, and the switch to TSMC 4N gave a massive improvement in efficiency. AMD did the big clock speed increase last generation, but otherwise CU * clocks seems to be relatively consistent as a performance estimator between RDNA2 and RDNA3.
RX 6800 XT = 72 * 2.25 = 162
RX 7800 XT = 60 * 2.43 = 145.8
So you can see that RDNA 3
does improve, but it's not by a lot. Some of that improvement is certainly due to the doubling of compute per CU as well, which in practice seems to improve gaming performance by maybe 15~20% compared to RDNA2.
RX 6650 XT = 32 * 2.635 = 84.32
RX 7600 = 32 * 2.655 = 84.96
In practice, probably partly because Navi 33 and Navi 23 are so similar, the gains here are even smaller. RX 7600 ends up maybe 5% faster than RX 6650 XT. That's probably again thanks to the doubling of compute, but it's interesting that the benefits here seem quite a bit smaller than with Navi 32 vs Navi 21 above.