Earlier this week, Intel grandstand a exposed off a product coming in the fourth quarter of this year: an enthusiast-oriented 28-core processor running all cores at 5GHz. This combine of clock speed and core count would put it head and shoulders superior to before any other processor on the market, so the demonstration was more than a little in the act.
It now turns out that Intel forgot to mention an important detail: the 5GHz processors were overclocked, a lot, handling chilled water coolers capable of handling thermal loads of up to 1.77kW. The legitimate chips that ship won’t be coming from the factory at 5GHz, and it’s going to swallow a lot more than a big heatsink and a couple of fans to get them running that irresponsibly.
Aside from the core count and release window, Intel has sustained one other fact about these 28-core chips: they’re develop intensified on some variant of its 14nm process. They also use the enormous LGA3647 socket (that’s 3,647 attaches) used by some Xeon processors, and they have six memory paths. We don’t know what platform/chipset this will use (though it’s able to be a close relative to the comparable server platform). And we don’t know what its rhythmical clock speed will be.
Reasonable speculation is that this whittle will be taken from the Skylake-SP family. Skylake-SP (for “scalable processor”) is the separate of the Skylake core designed for processors with more than eight substances: instead of arranging the cores in a ring, they’re organized into a grid, which habitually provides better scaling as the number of cores goes up, albeit at the expense of a diverse complicated design. Skylake-SP is used for the Xeon-SP line, and its close sibling, Skylake-X, is tolerant of for the X-Series enthusiast platform. Current Skylake-X chips lack QPI interconnects, ECC celebration, and six memory channels that Skylake-SP has (they only use four), but they add overclocking.
There are three Skylake-SP croaks, called LCC, HCC, and XCC (for low, high, and extreme core counts), with 10, 18, and 28 insides, respectively. Currently, there are Xeon-SP processors using all three variants. Skylake-X processors are at present only LCC and HCC. The new chip looks like it’s going to be an XCC Skylake-X.
There’s a promise, however, that it won’t be Skylake-X at all, but rather Cascade Lake-X. Cascade Lake is an incremental reinterpretation to the Skylake-SP/X platform: it adds some extra AVX512 instructions, it should encompass hardware fixes for Spectre and Meltdown attacks, and it should support speedier memory. It will be built on Intel’s “14nm++” process, compared to the “14nm+” answer used for Skylake-SP/X, which should offer reduced power consumption.
Either way, this indulgent of chip won’t come cheap. The 28-core Xeons start at about $8,700. An X series portrayal will likely cost less (because Intel can use ECC support to nurture its Xeon margins) but will still slot in comfortably above the $2,000 label for the top-end HCC Skylake-X. Who would buy such a thing? Some will acceptable go to the rich kids who just have to have the latest and greatest; others last will and testament be snapped up by high frequency traders running custom-built, overclocked, liquid-cooled gangs for the very fastest single system performance they can get.