Intel's MIC Compute Accelerators May Adopt QPI Bus Instead of PCI Express

Intel's Many-Core Accelerators May Adopt Common Server Infrastructure to Gain Market Presence

by Anton Shilov
11/15/2011 | 10:33 PM

Although Intel Corp. officially remains tight-lipped about the form-factor of its next-generation MIC compute co-processor code-named Knights Corner, it looks like the chip will use conventional land grid array (LGA) form-factor and will be plugged into sockets similar to those that are intended for Intel Xeon microprocessors.

Common PCI Express bus and slots allow compute accelerators for highly-parallel workloads from companies like Advanced Micro Devices or Nvidia Corp. to be easily installed into servers powered by Intel Corp.'s microprocessors, which greatly simplifies design of such machines, but also has its drawbacks. However, when it comes to the first commercial compute accelerators based on Intel's own many x86 integrated cores (MIC) architecture, the chip giant might have chosen an opposite approach and utilize proprietary Quick-Path interconnect (QPI) bus and infrastructure.


During the first public demonstration of KNC co-processors at a supercomputer event, Intel's Rajeeb Hazra, general manager of technical computing at Intel datacenter and connected systems group, showed off a chip presumably in BGA or LGA packaging, not a card. According to Intel, the Knights Corner accelerator has over 50 cores and delivers 1TFLOPS of double precision floating point performance, as measured by the double-precision, general matrix-matrix multiplication benchmark (DGEMM).

Even though PCI Express allows AMD and Nvidia to easily adopt their high-end graphics processing units (GPUs) for high-performance computing (HPC) needs, usage of PCIe means that a number of useful features (e.g., unified address space, quick access to data located in main CPU's memory, etc.) are either inefficient or hard to implement. In case accelerator is installed into similar sockets and uses common QPI bus, it can have the same features as central processing units (CPUs) and have access to the same data-sets and be more transparent to programs. In fact, many years ago AMD has already proposed so-called Torrenza multi-socket platform that allowed third-party accelerators to be installed into AMD sockets.

Installation of compute accelerators into typical sockets also allows to rather easily install more of such devices into servers or onto special expansion boards. While sockets do not allow usage of high-speed GDDR5 or alike memory, it is easy to install more memory per compute accelerator thanks to high amount of DDR3 memory slots per CPU socket than onto an add-on card. Furthermore, proprietary sockets mean that only Intel's own platforms can use MIC accelerators and it will be harder for AMD or Nvidia to sell their FireStream and Tesla cards for servers based on Intel's CPUs and MIC accelerators. Finally, it is cheaper for Intel to sell MIC accelerators in traditional LGA packages than build expansion cards like FireStream or Tesla.

It is unknown what socket type does the Knights Corner use, but usage of something like LGA2011 with quad-channel memory controller and plenty of power supply seems to be a logical choice.

Given the fact that KNC accelerators are not available and will be released sometimes next year, it is easy to expect Intel to change its approach and adopt a different form-factor for its CPUs.