Chipzilla has provided details of its Knights Landing, up-to-72-core Xeon Phi supercomputing chip.
Knights Landing be hitting the shops in 2015 and will use Intel’s 14nm process.
Significantly, Knights Landing will be a standalone processor, rather than a slot-in coprocessor that has to be paired with standard Xeon CPU. It will also have up to 16GB of DRAM 3D stacked on-package.
This will give it up to 500GB/sec of memory bandwidth, which is not to be sneezed at and Chipzilla tells us it will manage three teraflops per socket when it is going downhill and the wind is behind it.
This means that it should be easy to build a 100+ petaflop x86 supercomputer and with a bit of planning an exascale monster or two.
Intel currently runs the Xeon Phi, which is dubbed the Knights Corner. This is just a PCIe expansion board with an up-to-61-core Intel MIC chip. These cores are based on the original P54C Pentium core but with a lot of modern add-ons such as 64-bit support, 512-bit vector registers, go faster stripes and spoilers.
Knights Landing is a major rethink. The P54C cores are replaced with up to 72 Silvermont cores which can use AVX 3.1 instructions.
This means that the beast can manage six teraflops of single precision math, or three teraflops of double precision math without having to take off its shoes and socks and counting its toes.
Haswell’s head explodes when it has to do 500 gigaflops of double precision math, or one Italian tax return.
Knights Landing should manage between 14 and 16 gigaflops per watt which is better than the most efficient supercomputers currently max out at around 4 gigaflops per watt.