[rektide]

I can't wait to see a similar review of Bergamo/Zen 4c. Die shot analysis shows the cores themselves, not counting the L3 cache, are also 35% smaller than regular cores. But most of the specs look the same. There's a huge mystery here and there have got to be some interesting sacrifices.

Maybe maybe maybe 4c is chiefly a process tweak, that going lower power let them use more aggressive process scaling. Maybe. But that's still a huge die-space reduction that makes me think there's some secrets like here in this chip, that chipsandcheese.com was able to uncover with read deep prying/benchmarking.

I do wish the graphs in this article had a comparison processor or two. A regular Zen 3 would be great.

[toast0]

As I recall, from the Zen 4c articles, the big concessions are slower L1/L2, which allows fewer transistors per byte and so smaller area; lower clock targets, which allows less buffering to make the clock domains work and so smaller area. The L3 is about the same as a Zen 4 chiplet, which means half the size per core. They also dropped support for v-cache, which saves a little bit of area too.

Prior to Zen4c, server and desktop chiplets were the same, and desktop chips need to clock to the moon for competitive reasons, but server chips aren't going to go to really high clocks anyway; the thermals are too challenging, so if there's enough demand to justify a separate spin for servers, it makes a lot of sense to tweak the design for reasonable clock rates that will be seen there.

[touisteur]

I have had great success with the 'low-core / high-frequency' line-up in every step of the Zen+EPYC architectures.

They always have these (thanks to CCX and chiplets) SKUs and for latency-sensitive applications (or single-core non-parrallelized workloads), the high-frequency ones are amazing. You don't get the actual frequency/perf ratio, but they're relatively cheap (compared to the Intel SKUs with similar cores/freq) and damned reliable.

[redroyal]

Frequency optimized parts were created to sidestep software licensing based on core count. Often w/ higher cache. Like Milan-X

[phonon]

https://www.semianalysis.com/p/zen-4c-amds-response-to-hyper...