Anandtech has just published a large editorial detailing Intel's Knights Landing. Mostly, it is stuff that we already knew from previous announcements and leaks, such as one by VR-Zone from last November (which we reported on). Officially, few details were given back then, except that it would be available as either a PCIe-based add-in board or as a socketed, bootable, x86-compatible processor based on the Silvermont architecture. Its many cores, threads, and 512 bit registers are each pretty weak, compared to Haswell, for instance, but combine to about 3 TFLOPs of double precision performance.
Not enough graphs. Could use another 256…
The best way to imagine it is running a PC with a modern, Silvermont-based Atom processor — only with up to 288 processors listed in your Task Manager (72 actual cores with quad HyperThreading).
The main limitation of GPUs (and similar coprocessors), however, is memory bandwidth. GDDR5 is often the main bottleneck of compute performance and just about the first thing to be optimized. To compensate, Intel is packaging up-to 16GB of memory (stacked DRAM) on the chip, itself. This RAM is based on "Hybrid Memory Cube" (HMC), developed by Micron Technology, and supported by the Hybrid Memory Cube Consortium (HMCC). While the actual memory used in Knights Landing is derived from HMC, it uses a proprietary interface that is customized for Knights Landing. Its bandwidth is rated at around 500GB/s. For comparison, the NVIDIA GeForce Titan Black has 336.4GB/s of memory bandwidth.
Intel and Micron have worked together in the past. In 2006, the two companies formed "IM Flash" to produce the NAND flash for Intel and Crucial SSDs. Crucial is Micron's consumer-facing brand.
So the vision for Knights Landing seems to be the bridge between CPU-like architectures and GPU-like ones. For compute tasks, GPUs edge out CPUs by crunching through bundles of similar tasks at the same time, across many (hundreds of, thousands of) computing units. The difference with (at least socketed) Xeon Phi processors is that, unlike most GPUs, Intel does not rely upon APIs, such as OpenCL, and drivers to translate a handful of functions into bundles of GPU-specific machine language. Instead, especially if the Xeon Phi is your system's main processor, it will run standard, x86-based software. The software will just run slowly, unless it is capable of vectorizing itself and splitting across multiple threads. Obviously, OpenCL (and other APIs) would make this parallelization easy, by their host/kernel design, but it is apparently not required.
It is a cool way that Intel arrives at the same goal, based on their background. Especially when you mix-and-match Xeons and Xeon Phis on the same computer, it is a push toward heterogeneous computing — with a lot of specialized threads backing up a handful of strong ones. I just wonder if providing a more-direct method of programming will really help developers finally adopt massively parallel coding practices.
I mean, without even considering GPU compute, how efficient is most software at splitting into even two threads? Four threads? Eight threads? Can this help drive heterogeneous development? Or will this product simply try to appeal to those who are already considering it?