It's been some time since Intel introduced the Knight's Corner generation of its Xeon Phi co-processor, an add-in card that looks like a video card and acts as a floating point co-processor. It has slowly released the details, which made it clear what was taking so long. Now Intel has laid all of its cards on the table.
Intel held a Data Center Day event where it talked about multiple facets of its enterprise strategy and among them was Knights Landing, the successor to Knight's Corner.
For starters, it has a whole new core. The old Knight's Corner was based on Pentium cores. Yes, Pentium, the chip that shipped in 1995. The design was to tie the core with a high-speed interconnect and split up the work. The mesh connecting the cores has been reworked from the "ring" design of Knight's Landing to a 2D design that lets data move along an X and Y axis.
The new design features Silvermont cores, which are used in the Atom processor. Atom isn't exactly known for HPC performance but at least its architecture doesn't date back to the Clinton Administration.
Knights Landing has 36 tiles, each with two Silvermont cores and connected by a high-speed mesh that allows the tiles to talk to each other. Each tile has 1MB of L2 cache and each core can execute up to four hardware threads, which is better than a Xeon. It can only do two threads per core.
On top of it all is 16GB of on-package multi-channel DRAM (MCDRAM), which is Intel's answer to AMD's High Bandwidth Memory solution. It's meant to act as a very fast high speed interconnect between cores.
The two VPUs on each core is where this will be a true HPC device. Each one will come with a 512-bit AVX vector unit which will do all the heavy mathematical work on double-precision floating point. Intel is promising 3TFlops of double precision floating point, three times what the previous generation could do.
And because Knight's Landing has an actual modern Intel x86 core, the cards can now boot their own OS and run apps. Prior generations worked as co-processors and were dependent on a boot processor (a Xeon). Also, that eliminates the PCI Express bottleneck of having to go through the PCIe card to talk to the CPU for everything from the OS to apps.
Also because it's a x86 at its heart, Knights Landing is binary compatible with existing Xeon and Xeon Phi apps, no recompile needed.
Finally, the processors will communicate via an interconnect called Omni-Path Architecture. It addresses the problem of CPUs trying to communicate in a supercomputer system that can fill a basketball court.
Omni-Path supports four lanes per connection, and each connection has a 25 Gbps transfer rate. It works over both optical and copper cables and runs at the lowest layers of the network stack for very low latency.
Knight's Landing is expected later this year.