Erik P. DeBenedictis
Sandia National Laboratories
P. O. Box 5800 MS 1319
Albuquerque, NM 87112
(505) 284-4017 (o+m)
other contact data suppressed

Selected documents approved for public release as part of my work for Sandia National Laboratories


I've been encouraged by colleagues to post some documents to the Internet related to the development of the Cosmic Cube.

At Caltech in 1981, Eugene Brooks and I started a project that was called both the Homogeneous Machine (HM) and the Nearest Neighbor Concurrent Processor (NNCP). The project grew over time to include many people, as will be described shortly. The 64 processor version of this machine was subsequently renamed to "Cosmic Cube" and became a focal point for a growing number of people to develop parallel scientific codes. Users and Government funders followed the 3 MFLOPS Cosmic Cube with a series of exponentially larger and more powerful computers for 25 years, up to IBM's current 360 TFLOPS Blue Gene/L supercomputer at Livermore. This represents a 100 million fold scale-up!

The story I can contribute on this page is to give some background as to why the HM/NNCP/Cosmic Cube architecture survived repeated downselects and received 100 million fold scale-up where its competitors faltered.

Readers may draw their own conclusions from the original documents I've linked to this site, but my guidance is to look at two things:

  1. the strong connection between the architecture and semiconductor scaling and
  2. the presence of an user community with a scalable problem right from the time of the architecture.
Competing architectures of the time (SIMD, shared memory, vector), each fared less well on one of these criteria.

The early 1980s were a remarkable period at Caltech:

All these people were working together. Using unmodified microprocessors, HM/NNCP/Cosmic Cube was the full beneficiary of Moore's Law for microprocessor performance scaling over almost the entire existance of microprocessors.

Specifically observe that a key 1982 proposal document for the 3 MFLOPS HM/NNCP/Cosmic Cube contained a diagram (figure 1, page 3) and justification for scaling to a 500 nm semiconductor process and a system very much like the IBM Blue Gene system (noting that Blue Gene continued scaling another factor of four to 130 nm).

I have linked digitized versions of documents from the 1980s to the paragraphs below.

The Cosmic Cube project started in the summer of 1981. By the end of the summer, there was a propoosal document and a DOE report describing intentions as well as prototype hardware. ARPA funded the development of the initial hardware based on this technical plan, complete with a photograph of working prototype hardware and with systems software operational.

The four-processor, wire wrapped system became the basis for the paper Glueball Mass Calculations on an Array of Computers that included many of the simulation principles and computer systems analyses that were repeated many times in subsequent decades.

By October, 1982, the PC boards were being fabricated for the 64-processor hardware that was later to be called the Cosmic Cube.

However, I had graduated from Caltech and needed a job. I used this interview presentation and other presentation materials, and got a job at Bell Labs. As I left Caltech in December, 1982, I put the hardware lab notebook and the other hardware lab notebook away. 25 years later I scanned them and put on the Web.


Date of this document: January, 2008 (limited update March 2015).