Intel's slides for their new Xeons all focus heavily on trying to talk down Zen, using insane unfair comparisons, comparing one Xeon to a 1800x desktop chip clocked at 2.2Ghz, mentioning in every slide that EPYC uses a desktop die... while highlighting the coherent links that give multiple 50GB/s links from one die to another.... something literally not remotely used in desktop in any way shape or form and not even used in single core server, only in 2S systems. In previous years they wouldn't even be mentioning AMD, now they are mentioning EPYC in almost every slide and lying about it in almost every single slide.
An interesting thing is, Intel are critical of the L3 cache latency as being on average high, but the local L3 latency is extremely low while Intel's went with very low L3 causing hits to go out to memory with massively worse latency than AMDs local L3. So when software can keep the majority of data a cpu needs in local L3, then the latency is drastically lower than Intel chips with a neutered L3 who has to go out to memory more often as a result. They've upped L2 but had to dump a lot of L3 to fit everything on to one die. Again this is where the many die strategy works, Intel with a massive die on the 28 core can't afford more features, it had to cut back in multiple areas, AMD could afford to have more cache, more memory channels, more pci-e and more core precisely because adding 25mm^2 of stuff to a 170mm^2 die to add those features is easy where as adding 4x 25mm^2 worth of features to a single 680mm^2 die taking it up to a 780mm^2 die would have been the difference between getting 10 working cpus per wafer and 1 working cpu per wafer and 10 would already be a disastrously bad yield at 680mm^2.
For one of the main things they criticise EPYC for, L3 latency, real world worst case latency for the L3 won't happen often while Intel having just over half as much will be a problem much much more often.
Semiaccurate saying they have confirmation that Intel's way of chopping up the chips with neutered features has cost them customers who will go with EPYC instead, so great move by Intel... for AMD that is, it's bad for Intel. The whole launch seems terrible from Intel. Why are they even saying AMD just reused a desktop die... if they reused a desktop die and provided better server features with EPYC what exactly does that say about Intel who spend billions in R&D for server chips... they made something dedicated and it's a bit faster in some things, slower in others and has some significant drawbacks, than a chip AMD threw together from desktop parts? That just makes Zen sound amazing and Intel sound stupid. That messaging along with angering customers by artificially neutering features and demanding literally thousands extra just to not cripple a feature like memory capacity, it's just crazy.
The messaging of, hey, AMD had 5 years to work on this one architecture and we've worked on this for a year since Broadwell-e and look at these benchmarks where we still win sounds so much better but they went for the absolutely worse message possible... AMD barely even tried and basically matched us.
