Good review, thanks!
Looks exactly like I suspected.
The multi CCD penalty in gaming is still real, so it consistently performs lower than the 9800x3d in games, but on average the penalty seems less than in the past, maybe because of slightly higher clocks.
It would be interesting to see how it performs in games with one of the CCD's disabled.
I wonder how much better a 9950x3d with one CCD disabled would perform compared to the 9800x3d...
Based on the clock speeds (5.7Ghz vs. 5.2Ghz) the upper bound should be 9.6% at stock clocks, but in reality it will likely be lower, as RAM bandwidth/latency will likely hold it back.
Lets not forget that overclocking is now possible with x3D CPU's. I'm going to presume that the CCD's in the 9950x3d are higher bins to hit the 5.7Ghz clocks, but the question to me is this:
If you are a gaming extremist and overclocker, and want the best gaming performance bar none....
Is it better to buy a 9950x3d for $699 (compared to the $479 9800x3d) and disable one CCD (to get rid of the CCD penalty)?
Or, if you are going to overclock both anyway, are you likely to be limited at about the same clocks/performance.
In other words, can you get higher by overclocking a one CCD 9950x3d than by overclocking a 9800x3d.
I've already seen many 9800x3d overclocks it 5.7Ghz, so we know it can tie the stock 9950x3d, but the question is, is that 9950x3d right at its limit, or will that higher bin allow it to go even higher?
...and if it does go even higher, (6+Ghz?) will it even matter, or will AMD's fabric/memory controller design hold it back anyway?
This would be a really cool test, especially if a few different memory configurations are thrown in for good measure. Which would be faster for games on an AMD system when really pushing the clocks like this? DDR5-6400 in a 1:1 MCLK to UCLK configuration, or DDR5-8000 in a 2:1 MCLK to UCLK configuration.
The DDR5-8000 will get slightly less top level memory bandwidth due to a slightly lower UCLK, but on the flip side, with the RAM running at 4000Mhz (8000MT/s) penalties from things like refreshes are going to be much smaller, as those don't use memory controller bandwidth, and instead operate at the speed of the RAM. Of course, timings and submitting will likely make a difference too.
I have yet to see a really good review that combines all of these complexities of overclocking X3d chips, multiple CCD's and RAM settings when pushed to the limit.
