3 comments

[ 5.4 ms ] story [ 15.5 ms ] thread
So, they're effectively killing off AVX-512 in the consumer space before it even really caught on. Makes me wonder whether AMD will bother with it at all. Either way, given the QoL improvements that were in the instruction set, this is very sad news for those of us writing SIMD code on a regular basis. Guess I'll be writing AVX-2 for another decade...
Very curious about the performance gains being hinted here: 19% for Golden Cove (vs. Cypress Cove), and 1.5X for Gracemont (vs. Skylake Atom?).

The 19% Golden Cove jump doesn't seem very impressive given Cypress Cove was a regression. As noted, it should've been measured against the newer Willow Cove (e.g. Tigerlake cores) for better indication.

As for Gracemont, 1.5X perf sounds great but how capable will be it for everyday tasks? E.g. if the Big/Performance core has to spin up for almost everything, then it won't be very useful. Furthermore Skylake was initially released in 2015 so 1.5X better after 6 years is way below Moore's Law.

Skylake is not an Atom. Gracemont is post-Tremont, the current generation of Atom. Tremont and Gracemont have several interesting properties for systems programmers, including UMWAIT which allows a user-mode program to wait for a write to a given address, and a huge shared L2 cache that all cores can access in 17 cycles. The former have apparently been ported to the new Cove high-performance cores, which will unlock some really neat abilities.

I've long wanted a consumer Tremont part, without any of that Cove junk, because I think it is really offering something unique to the programmer. If I have to buy a hybrid CPU to get it, I guess that's fine.