Want to know what is hilarious about upcoming Intel lineup with X299 Chipset?
It was specifically designed for past 3 years for Socket 2066, and they kept in mind to offer only 12 cores. When Intel heard AMD will offer 16 core CPUs, they decided to add 14, 16 and 18 core CPUs.
The chipset (PCH I/O ) and the cores don't really have all that much to do with each other. They have to do a proper handshake on boot, but past some 'uncore' coupling the span isn't all that.
The thing is... Intel might be forced to bring another socket for those CPUs, Socket 2066 v2, because the 14, 16 and 18 core CPUs might not be compatible with Socket 2066.
Forced? Intel slapped a chip die for an LGA 1151 socket into the LGA 2066 package. The extra PCI-e and memory pins that the small chip couldn't drive are left dangling. So there the chips isn't uniform in memory and PCI-e coverage. As mentioned above on this page some of these 2066 packages have PCI-e lanes kneecapped off ( only 28 lanes of the 44 enabled. ).
Given the above it won't be all that hard to kneecap a LGA 3640 intended die. Kill off the Omnipth pins. Kill off the 10GbE support pins. Kill off 2 of the 6 channel memory controller paths.
It won't be as fast as the Gold and Platinum variants on diverse highly parallel workloads, but it will probably be substantively cheaper kneecapped.
The cheap path would be just to kneecap the stuff present by not hooking it up. A more refined path would be to just take the x86 cores and L3 subsystem and wrap a 2066 'uncore" I/O package around it. I suspect it is much cheaper and quicker just to turn some of it off and/or not connect it.
[ this article discusses there are two "core" like sections on the die that don't look like the other cores. So have 18 cores instead of what supercially looks like 20.
http://www.anandtech.com/show/11464...ng-18core-hcc-silicon-to-consumers-for-1999/2
If those are OmniPath/FGPA/10GbE components that leverage the extra pin bandwidth out of the larger socket they Intel can simply flip those off. They are going to need TDP headroom since they are going to have to clock these higher than what they were primarily design for.
Intel has stuffed a core reduced HCC die into a Xeon E5 1680 before. second QPI links weren't used. ]
The only thing Intel is being "forced" to do here is lower their prices.
Xeon E5 v4 18 core with a "higher Turbo".... $2700
http://ark.intel.com/products/91755/Intel-Xeon-Processor-E5-2697-v4-45M-Cache-2_30-GHz
At Anandtech link above Core i9 7980XE ... priced $1999. About a $700 savings. I think Intel's Gold/Platinum 18 cores will be higher than that but they probably will kneecap a variant so that can hit lower price points. ( that is exactly what they have done with the lower end of the i7-78xxX line up. )
For the Skylake-W line up Intel could use one or two of those HCC versions for a Xeon E5 1690 1695 (for instance ... I suspect Intel is going to come up with a much more goofier names.... "Corecrusher" versus "threadripper". .... like Gamera versus Godzilla ) that would correspond to the 14 and 16 versions. Personally, I'd leave the 18 core variant alone. They could add an "upclocked" version to the Xeon Sliver line up and keep the LGA 3640 slot to feed the cores with memory bandwidth. I think the 7980XE is going to work much better as a ego processor that is run most of the day in Turbo 3.0 mode on a couple of cores and largely has some geekbench bragging rights the rest of the day. ( I think 18 higher clocked cores with just 4 memory controllers is going to run into constraints. Not crippling but on the diminishing returns curve. )
If someone wants 16+ cores with high RAM with ECC just bump them up to the Bronze/Sliver/Gold/Platinum zone. For the never enough cores crowd that will be a better solution than the single socket solution. Chasing them with Xeon E5 1600's has limits. Intel really only needs to match the AMD core count of 16 for the folks who aren't primarily focused on hot rodding and memory errors don't matter.
