[Beowulf] Power8

[Bill Broadley]

Sounds like a potentially interesting CPU/platform for HPC.  Of
particular interest:
1) similar quad socket performance to intel's best
2) embracing 3rd parties access to cc memory
3) up to 8 off chip memory controllers with cache (centaur chip)
4) allowing 3rd party motherboards
5) IBM exploring getting out of fab business
6) IBM apparently open to license various aspects of the power 8
7) substantially better off chip memory bandwidth "230GB/sec sustained"

Open questions:
* is the interconnect cache coherent?
* does a chip <-> chip interconnect come at the cost of one of the
  memory channels/centaur chips?
* latency of interconnect?
* will 3rd parties like nvidia be able to do something compelling with
  cache coherent access to main memory?
* will 3rd parties be able to replace the dram <-> centaur chip
  with something that provides more performance?  Seems like one of the
  dram fabs could tune the bandwidth/latency/power, ddr3/4/5/gddr
  chips, and number of outstanding memory transactions.  Much like
  current SSDs and GPUs do.
* How much will a barebones dual socket OpenPower system cost?
* Can OpenPower push Intel on performance per $ or performance per watt
* Is tunneling CAPI over PCI-e sane?
* Will anyone smaller Google, Facebook, and Amazon be able to buy
  get good pricing?


Information I've found so far:

Motherboard: https://plus.google.com/u/0/111282580643669107165/posts

For a S824 power8 @ 3.5 GHz, 4 chips, 6 cores/chip, 8 threads/core:
 1370 SPECfp_rate2006
 1750 SPECint_rate2006

Interesting "48-way drawer" with 2-hop interconnect, page 24.  Each chip
has 3 38.4 GB/sec links and 3 12.8GB/sec links to neighbors:
  http://alcoidstein.de/IBM%20Intro%20to%20POWER8%20Processor.pdf

Overview:
http://www.extremetech.com/computing/181102-ibm-power8-openpower-x86-server-monopoly