Cray-3 Range

The Cray-3 and Cray-4 prototype are the last in an impressive line of computers designed by Seymour Cray. Similar to the Cray-2 in size and cooling arrangements but the modules shrunk and the technology changed from silicon to gallium arsenide

Cray-3 memories by Steve Gombosi

From a comp.unix.cray posting

Graywolf (“S5”) was installed at NCAR. Like all NCAR supercomputers, until fairly recently, it was named after a Colorado locale. This was the only Cray-3 shipment, installed in May 1993, the machine was a 4-processor, 128 Megaword system. Two problems in the Cray-3 system were uncovered as a result of running NCAR’s production climate codes (particularly MM5): a problem with the “D” module causing intermittent problems with parallel codes, and an error in the implementation of the square root approximation algorithm which caused incorrect results for certain data patterns (kinda like the Pentium divide bug 😉 ). These were rectified and replacement CPU modules were installed, although I can’t remember the date.

The machine ran NCAR production until CCC folded in March, 1995. Since NCAR never paid for it, at some point we reduced the CPU count to 2 and let the machine run essentially unattended. I’m not too sure when that happened, although it marked the end of my regular commuting between Colorado Springs and Boulder.

There were a total of 7 Cray-3 “tanks” constructed. S1-S4 were single “octant” tanks (the smallest that could be constructed) which accomodated up to a 2 processor/128MW configuration. S5 and S6 were two-octant tanks. S7 was a four-octant tank which we used as a software development and benchmarking platform. S6 was chiefly used for system testing.

S1-S3 were diverted to Cray-4 testing once the Cray-4 project built up steam. S4 was diverted to the quite possibly suicidal Cray-3/SSS project after S7 became available (S4 was previously our software development machine).

For those of you who have Cray-3 posters lying around (by the way, I took all the photos on that poster as well as the Cray-3 and Cray-4 brochures and all the annual reports except the first two):

1) The big photo is of S5
2) Seymour is leaning on S5 (and you have no idea how hard it was to get him to hold still that long while wearing a suit…or to talk him into that particular pose)
3) The two “cooling system” photos are S6
4) The hand holding the module is mine 😉

Cray-3 modules were 4x4x0.25 inches in size. Each module consisted of a multi-layer “sandwich” of PC boards (69 electrical layers), with 2 layers of 16 1×1 inch stacks. The stacks were the circuit boards containing the actual circuits (GaAs for logic, SRAM for memory modules). There were 16 bare GaAs chips mounted to each side of a logic stack. I think there were 12 bare SRAM chips on each side of a memory stack (the logic chips were square, the memory chips were rectangular).

Cray Computer Corporation From old link

Cray Computer Corporation Status: Not active. Overview of Organization:

In 1985, as soon as the Cray-2 had reached production, Seymour Cray and team started work on the Cray-3, which was to be a GaAs implementation of the Cray-2. It was thought that GaAs technology would be viable by 1988/89, which was the original production target date for the new machine. However, the Cray-3 is still not available and is not expected until 1991 at the earliest.
Some of this delay is due to problems with using advanced technology, especially in the area of packaging, and some is due to the physical moves the project has made. In the spring of 1988, the Cray-3 project moved from Chippewa Falls to Colorado Springs. This was done to focus the project on the Cray-3, since Chippewa Falls was mainly geared to producing the Y-MP. In November 1989, the Cray Research Inc. (CRI) management decided that it could no longer support both the C90 and Cray-3 projects. The most promising project was the C90, so to avoid closing down Seymour Cray’s own project, the development of the Cray-3 was spun off into a new company called Cray Computer Corporation (CCC). Cray Research transferred a facility and other equipment to CCC, along with an $85 million promissory note to fund future development of the Cray-3. Cray Research hold 10% of the stock of CCC, the other 90% of stock being distributed among Cray Research stockholders. This has prompted another change in location for the Cray-3 project and caused more delays.
CRI dealt with Rockwell for the GaAs chips for the original design of the Cray-2. By the time the Cray-3 was being designed a small spin-off company from Rockwell called Gigabit had started up; they now supply the Cray-3 project with chips.
In November 1988, at a supercomputing conference in Florida, Seymour Cray described the architecture, technology and packaging of the Cray-3, and some design targets for the Cray-4. The Cray-3 will be 12 times faster than the Cray-2. A factor of three speedup will come from the use of GaAs technology, which will give the Cray-3 a 2 ns clock cycle. The remainder will come from using four times as many processors as in the Cray-2. Seymour Cray noted that one problem with the Cray-2 was its primitive interprocessor communication mechanism. As a result, the Cray-3 will use the communication architecture of the Y-MP series. The expected performance of the Cray-3 will be 8000 MIPS or 16 GFLOPS. it will have 512 Mword of main memory.
The entire machine will fit inside a cuboid cabinet measuring 1′ on each side. The reason for the small size is that reducing the machine’s linear dimensions by a factor of three gives a corresponding improvement in speed, but a reduction in volume of 27. The maximum allowable wire length throughout the machine is one foot. Producing a machine of such tiny proportions requires radical packaging technology, an area in which Seymour Cray excels. Each processor will be constructed from four modules. The whole machine, including the central memory, will be built from 200 modules, each contain 1024 VLSI circuits, and be built from circuit wafers stacked and bonded together. The machine will be liquid cooled using technology developed for the Cray-2.
The largest problem faced on the Cray-3 project is how to automate machine production. There are many connections to be made in using a complex interconnection scheme. To do this, laser welding and cutting equipment is required, as is robot assembly technology. Seymour Cray is already thinking about the Cray-4. This is likely to have a 1 ns dock and will test the limits of GaAs technology.
After the Cray-4, he expects that a new material will be needed, possibly indium phosphate layers on a silicon substrate. “Cray-3 to have up to 8 dual-CPUs (i.e., a maximum of 16 vector processors), connected to a maximum of 4GB of central memory, with an interprocessor communication mechanism architecture based on the Cray Y-MP”; “One-foot cube for entire [Cray-3] machine, immersed in inert liquid fluorocarbon for optimum cooling”;

NCAR history page for Greywolf

“The Cray-3 was the brainchild of Seymour Cray. He had begun its development while still chief architect with Cray Research, Inc. (CRI). CRI had spun Seymour and the Cray-3 development team, located in Colorado Springs, CO, as an independent company in 1989, and Seymour named this new company Cray Computer Corporation (CCC). In the early, 1990’s, CCC had built four prototype Cray-3 systems and was building four additional systems which it hoped to market.

On May 24, 1993, CCC delivered a Cray-3 supercomputer to NCAR as a test and evaluation system. The gray, four-foot tall machine was called Graywolf, following an NCAR tradition of naming computers after 14,000-foot peaks in the Colorado Rockies. Seymour had decided to loan Graywolf to NCAR, in part, to expose it to use by NCAR scientists and thus help CCC to develop its operating system, compilers and libraries in a real-world environment. In doing so, he had said “If the Cray-3 can survive NCAR, it can survive anywhere.”

Graywolf was S/N 5 of the Cray-3 product line and the first ‘production-ready system. The first four systems had been built and used in-house at CCC for hardware design and software check-out. The Cray-3 was unique in that it used gallium-arsenide circuits, which could be clocked faster than silicon-based integrated circuitry at that time. The Cray-3 was designed to have eight “octants” of circuit boards, immersed in a circulating Fluorinert bath to keep the circuitry cool. Graywolf was a two-octant Cray-3 and, while capable of having up to four processors, it had two processors, 128 megawords (1 gigabyte) of memory, 20 gigabytes of disk space, and a clock speed of 2.08 nanoseconds (480 MHz) — the fastest clock speed of any supercomputer then available. It was connected via HIPPI to NCAR’s Mass Storage System. Graywolf ran CCC’s Colorado Springs Operating System (CSOS), which was based upon the Cray-2 version of Cray Research, Inc’s UNICOS operating system, CCC’s vectorizing Fortran and C compilers, optimized math libraries, debugger and performance tools.

The Cray-3 used gallium arsenide integrated circuits in place of silicon for all its logic circuitry. This circuitry and the system’s memory was contained under a translucent smoky gray acrylic lid in the top eight inches of the system cabinet. The larger portion of the Cray-3 cabinet beneath the lid contained the Fluorinert circulating pumps, power supplies, and system control and monitoring devices. Graywolf required 90 kilowatts of power and produced 310,000 British thermal units of heat per hour — enough to warm six 2,000-square-foot homes.

NCAR used Graywolf to run atmospheric and oceanic circulation simulations and was used for CCC software development. It is noteworthy that NCAR’s atmospheric circulation model helped CCC to discover a Boolean logic error in the square-root hardware of the Cray-3. The Cray-3 had introduced a new square-root instruction not previously available in the Cray-2 or Cray-1, and the hardware produced the wrong result for one out of every 64k operands. Once discovered, CCC provided a temporary fix in the compiler (i.e. don’t use the new square-root hardware) and provided a fix to the hardware three months later.

Graywolf was decommissioned on March 26, 1995, the day after CCC filled Chapter 11 bankruptcy, and was returned to CCC’s Colorado Springs facility.”

What might have been ...

Seymour and Cray Computer Corp was working on the Cray-4 and Steve reports “If we’d stayed in business another month or so we would have sold and shipped a couple of Cray-4s and things would have turned out very differently”.

CCC also did work on the Cray-3 SSS prototype, some special hardware with the special community but that did not result in a completed sale.

Also noted ….

The Cray-3 wasn’t the last of the CCC GaAs machines. There are probably two more which ought to be mentioned:

1) The Cray-3 SSS (Super Scalable System), which combined two Cray-3 processors with 65,536 PIM (Processor In Memory) chips.

2) The Cray-4. While that system never shipped to a customer, we had working systems in-house and running customer benchmarks when we went Chapter 11.

Scroll to Top