Discussion in 'Article Discussion' started by Gareth Halfacree, 8 Mar 2017.
Up to two Naples chips per server.
Should that be "now"?
Knowing next the nothing about server products how does, on the surface, Naples compare with Xeon?
Yes. Yes, it should.
We won't know until somebody gets to test 'em properly. Naturally, AMD says Naples is better - but it's going to run into the issue of compilers being heavily optimised for Intel microarchitectures, the lack of greater-than-two-socket support (though to be fair this is somewhat mitigated by each socket having so many cores and memory channels), and convincing companies to actually build Naples servers people can buy. It looks promising, though, and if Naples undercuts Xeon like Ryzen undercuts Core then it has a fighting chance of succeeding.
Would be wise to wait for the benchmarks on this one, though Ryzen has a lot of cores for example, it only supports AVX-128 instructions while Intel is at AVX-512.
Sorry when i said how does, on the surface, Naples compare with Xeon i should have been more specific, i should have said how does the core count, amount of I/O bandwidth (PCI-e lanes) and supported RAM compare.
Xeon tops out at 24 cores vs the 32 of Naples, 40 pci-e lanes vs 64 per socket (AMD says 128, but 64 of them vanish in multi socket config), quad channel vs octo channel and the 24 core chip costs over $8k vs most likely less for the 32 core chip.
So on first glance Xeon gets beaten to a bloody pulp.
But of course nothing is that simple, Xeons can go beyond two sockets, software has been optimized for Intel due to AMD having been gone from the scene for years, so Intel may still have an advantage due to that, plus there are specialized products that will make minced meat out of either one (Xeon and Naples are just the fastest all purpose chips), depending on workload you may have consider stuff like Xeon Phi, Nvidia Tesla or the upcoming Radeon Instinct and potential compatibility issues / bottlenecks etc...
I think AMD are onto the right track, technically speaking, but that won't be enough.
Intel will own a few compute-heavy scenarios no doubt, but there's more noise there in the form of non-x86 there too and the biggest chunk of the market is general purpose virtual compute. Scaling beyond two sockets there is largely irrelevant, where environments are almost overwhelmingly memory constrained. Even scaling to more than one socket is of questionable value when you're licensing by the socket, and probably aren't even touching the sides with a single 20+ core CPU.
In the end, mainstream success or failure is more down to partnerships than the competency of the platform...
Then there's the new totally-not-E7-any-more-guys Xeon 'Gold' and 'Platinum' bunch-o-cores chips coming soon to drop into LGA3647.
Separate names with a comma.