最近学习Intel Fortran编译器的自动矢量化和OpenMP多线程，看到Xeon Phi这个东西。矢量化和多线程在Xeon Phi上可以得到最好的应用，部分常用程序（计算量限制Compute-Bound）可以得到很大的加速比。遂起意自己购买Xeon Phi做练习之用。
在网上查了很多资料，Intel没有给出很完备具体的教程，大部分支持工作都是通过博客文章和论坛完成。
经过搜索整理，Xeon Phi的常用重要文档如下：
1，Intel将Xeon Phi的系统驱动、工具软件集和内部子系统等软件打包，称为MPSS（Manycore Platform Software Stack）。
各版本信息和下载见博客文章：
https://software.intel.com/en-us/articles/intel-manycore-platform-software-stack-mpss
2，Xeon Phi的硬件基本要求：
https://software.intel.com/en-us/articles/which-systems-support-the-intel-xeon-phi-coprocessor
-
- Double-wide x16 PCI Express slot for each coprocessor in the system
- BIOS support for memory mapped I/O address ranges above 4GB
- Up to 300W power delivery and sufficient cooling (varies by SKU)
3，常见支持主板：
https://streamcomputing.eu/blog/2015-08-01/xeon-phi-knights-corner-compatible-motherboards/
Gigabyte
Below are three motherboards Gigabyte suggested.
Only have one CPU to spare and chose a dual CPU board? Then “some PCIe, memory, network or I/O functions might be unavailable” – probably because there are less PCIe lanes. At least it works, when you want to do a relaxed upgrade.
Big plus for Gigabyte is the space they have for a double GPU for the last slot.
ASRock
In short: all “WS” motherboards. You will find on forums that the cheaper “X99 WS” won’t work and needs a special bios, but I asked and it seems to work out-of-the-box (untested). The WS-motherboards are on the whole price-range between €300 and €800, having all kinds of different options.
If you want 88 lanes with a single CPU, try the X99 WS-E10G with a 44 lane PCIe-switch. If you want a cheaper option with extra lanes, keep searching for other motherboards having the “PLX PEX 8747” controller or alike.
ASRock themselves suggest you use the X99 WS as a started motherboard, and then see what is needed for when you upgrade.
ASUS
ASUS seems to have the most info, with special “AVL lists” for each motherboard. So using Google, I could create the below list. As ASUS has very complete information, it’s easy to double check before you buy.
Motherboards with LGA 2011-3:
-
- Z10PA-D8: 2 x LGA 2011-3, 7x PCIe (x16x16)
- Z10PE-D16 WS: 2x LGA 2011-3, x PCIe (x16x16), 4 way x16
- Z10PE-D8 WS: x LGA 2011-3, x PCIe (x16x16), 4 way x16
Motherboards with LGA 2011:
-
- Z9PE-D8 WS: 2x LGA 2011-3, 7x PCIe (x16x8x16x8x16x8x16)
- Z9PE-D16-10G: 2x LGA 2011-3, 5x PCIe (x2x16x8x16x16)
- P9X79 WS: x LGA 2011-3, 7x PCIe (x16x8x16x8x16x8x16)
- P9X79-E WS: x LGA 2011-3, 7x PCIe (x16x8x16x8x16x8x16), 4 way x16
从性价比角度，我最终选择了Intel Xeon Phi 31S1P（57核，1.1GHz，8G内存）。决定先在主板Z9PE D16C/2L上测试。