The new overclocking features for AMD's upcoming Zen 3 Ryzen 4000 CPUs include a Curve Optimizer and the Infinity Fabric dividers, and we also get a confirmation that the new CPU lineup will feature 10-core SKUs, as AMD has upped the number of cores per CCX to 8.
https://www.notebookcheck.net/AMD-s-Zen-3-Ryzen-4000-desktop-CPUs-will-include-advanced-new-overclocking-features-and-will-also-come-in-10-core-20-thread-flavors.492432.0.html
That's odd.
As far we know, AMD will be using N7P node which doesn't have any density improvements over existing 7nm but only features 10% increased efficiency or 7% improved performance.
Also, we were told before not to expect core increases from Zen 3 (probably because there are no density improvements on the node they are using).
However, if AMD uses N7+ node which uses EUV, then they could gain 20% higher density... which would indeed allow AMD to transform their 8c/16th CPU's into 10c/20th CPU's on Zen 3.
Otherwise, I just don't see how AMD plans on putting more cores onto the same node without overstepping their TDP budget.
Similarly, for the mobile APU's, it would be better if they added more graphics cores to the chip as the CPU cores would experience a large leap in IPC anyway.
I agree. I think this would be an odd move. They already have a significant "cores" lead over Intel which does not look to change soon. I also think this would complicate Epyc designs. I think the big improvement is going to be 8 cores per CCX, cache and latency improvements plus the process improvements going from N7 to N7P which is IP compatible. N7+ is not IP compatible and there is not enough performance improvement between N7P and N7+ to justify that significant of a redesign.
Quote from: deksman2 on September 08, 2020, 17:16:23
That's odd.
...
However, if AMD uses N7+ node which uses EUV, then they could gain 20% higher density... which would indeed allow AMD to transform their 8c/16th CPU's into 10c/20th CPU's on Zen 3.
Otherwise, I just don't see how AMD plans on putting more cores onto the same node without overstepping their TDP budget.
...
I'm not 100% sure on what you are finding "odd". I don't think there was any expectation for having more cores than 16, which current zen already has.
The CCXs will have 8 cores instead of four, and have 32mb of cache instead of 16. This will reduce latency, It also potentially allows for an 8 core processor to have just one CCD, which will also improve latency.
A ten core part would have 2 CCDs with 5 active cores each - this may help improve usable yields.
Its all making perfect sense to me ... they are decreasing the modularity to take advantage of the "advantage" that monolithic designs have (low latency), while at the same time not really losing any advantage that their modular design allows. It seems an intelligent and natural progression.
Quote from: JimboJ on September 08, 2020, 18:28:38
Its all making perfect sense to me ... they are decreasing the modularity to take advantage of the "advantage" that monolithic designs have (low latency), while at the same time not really losing any advantage that their modular design allows. It seems an intelligent and natural progression.
Yup, that's what I was reading from this too.
The 10c/20t model would make perfect sense for the 4800x. The minimal difference between the 3700x and 3800x never seemed right to me compared to their other models. Now you could have a core increase for every tier.
Except marketing does not drive engineering....thankfully.