ETA Systems

ETA Systems was a supercomputer company spun-off from Control Data Corporation (CDC) in the early 1980s in order to regain a footing in the supercomputer business. They successfully delivered an excellent machine, the ETA-10, but lost money continually while doing so. CDC management eventually gave up and folded the company.

Seymour Cray left CDC in the arly 1970s when they refused to continue funding of his CDC 8600 project. Instead they continued with the CDC STAR-100 while Cray went off to build the Cray-1. Cray's machine was much faster than the STAR, and soon CDC found itself pushed out of the supercomputing market.

William Norris was convinced the only way to regain a foothold would be to spin off a division that would be free from management proding. In order to regain some of the small-team flexibility that seemed essential to progress in the field, ETA was created in 1983 with the mandate to build a 10GFLOPS machine by 1986.

ETA had only one product, the ETA-10. It was essentially a modernized version of the CDC Cyber-205 computer, and deliberately kept compatible with it. Like the Cyber series, the ETA-10 did not use vector registers as in the Cray machines, but instead used pipelined memory operations to a high-bandwidth main memory. The basic layout was a shared-memory multiprocessor with up to 8 CPUs (and up to 16 I/O processors), each capable of 4 double-precision or 8 single-precision operations per clock cycle.

Each CPU had up to 32MB of private memory, and the set of CPUs had common access to up to 2GB of shared memory. Most CPU load/store instructions could not touch the shared memory, however, and I/O operations could not touch the private memory, creating store-and-forward problems for the operating system. Fortunately, there were also interprocessor message buffers that lived outside the main memory space.

The main reason for the ETA-10's speed was the use of a liquid nitrogen cooling in some models to cool the logic components. Even though it was based on then-current CMOS technologies, the cooling allowed the CPU's to operate on a ~7ns cycle, so a fully-loaded ETA-10 was capable of about 4.5 GFLOPS. The design goal had been 10 GFLOPS, so the design was technically a failure. Two LN2-cooled models were designated ETA-10F and ETA-10G. Two slower, lower-cost air-cooled versions, the ETA-10Q and ETA-10P (code named "Piper") were also marketed.

Software was largely a disaster. When they intially shipped in 1986 there was no operating system for the machines. Programs had to be loaded one at a time from an attached Apollo Computers workstation, run, and then rebooted to run the next. At the time Unix was making major inroads into the supercomputing fields, but ETA decided to write their own EOS which wasn't ready when the machines were. A Unix-based system became available in 1988, at which point it looked like the machine might finally succeed. Many sites that had refused to pay for their machines due to the low quality of EOS found ETA System V completely usable and were willing to accept delivery.

In April 1989 CDC decided to fold ETA back into CDC, by which point 7 liquid-cooled and 27 air-cooled machines had been sold. At this point ETA had the best price/performance ratio of any supercomputer on the market, and its initial software problems appeared to be finally sorted out. Nevertheless, shortly thereafter CDC exited the supercomputer market entirely, giving away remaining ETA machines free to high schools.

External link

• Computer history TNO-FEL: the ETA10-P (http://www.tno.nl/instit/fel/museum/computer/en/eta10p.html)
• The ETA Saga (http://yarchive.net/comp/eta_peglar.html)