the GCOS 64

the GCOS 64 / GCOS 7 system

Hardware

The Level 64 has been jointly designed by engineers from Honeywell-Bull laboratory in Paris, from Honeywell in Billerica with a small Japanese participation from NEC.

This product line was originally thought in 1970 as "the" Honeywell New Product Line.
In 1973, its scope was reduced to medium systems designed and built by the French subsidiary of Honeywell.
NEC, an Honeywell licensee since the 1960s, maintained the initial objectives and developed its own Level 64's derived systems -the ACOS4 product line-.

Considering only the products designed by Bull and ignoring momentarily the tentative and aborted projects , three genuine processors have been successively designed:.

in 1974 the original processor code named P7,
in 1981 a relatively more powerful processor code named Leo or P7G
and in 1993 a new design code named Auriga .

In the 1990s, it was decided to discontinue the design of proprietary processors of DPS-7 product line and to implement the DP-7000 interior decor to Intel manufactured processors, initially on IA32 Xeon, later on IA64 processors.
That project was developed since 1997 under the code name of Diane. It was publicly disclosed in mid-2001.

Those basic processors were implemented in several technologies that allowed (or required) not negligible design changes. In addition, each technical model was delivered in several options leading to different marketing names. Price changes, to match the IBM competition, lead also to new models with sometimes artificial modifications. So, the marketing models identification is somewhat confusing. The initial model was named Level 64. From 1980 to 1987, all models were named DPS-7, and subsequently DPS-7000.

The P7 design was based around a 32-bits data bus and a microprogrammed processor. The hardware architecture used a relatively long word and microprograms were fetched from a read-only PROM control store. The same processor circuits was used to interpret the software "decor" instruction set and to handle interrupts from the channel devices. Software visible registers were implemented either in hardware registers or in main memory. A small (but at that time quite expensive) associative memory contained the address translation buffers between the segmented software address and the physical address in main memory. The main memory (max 1 M Bytes) was implemented with 4 Kbits semiconductor chips ( quite new for main frames in the early 70s ). The basic CPU (and peripheral processors) bipolar technology was Texas Instruments 74N small-scale integration chips with a few MSI circuits such as multi-bits multiplexers.

P7 was designed in "worst case" design with a guaranteed cycle of 2 MHz and was introduced as two slowed down marketing models the Series 60 64/40 and the 64/50. Tuning the processor on the longer data paths allowed to increase the frequency to 2.5MHz and to introduce a full-speed version of P7, the 64/60.

A new model called 64/DPS as Distributed Processor System (a marketing name gimmick used by all the Honeywell system without any significant technical relevance) was introduced in 1979. 64/DPS was based on P7. It included a larger (4MB) main memory and repackaged peripheral controllers reducing the number of frames of the central unit cabinet. The use of compatible 74S chips instead of 74N allowed to decrease the cycle time to 300ns. The reduced price of 64/DPS causes a significant increase of sales and even shortages in the production line.

the 64-DPS system

a P7G micropack

Around 1977, CII-HB decided to use the CML micro packaging technology for the P7G (and also for the DPS-88 in Phoenix). As the Angers plant capacity exceeded the P7G requirements, it was decided to use the same CML technology and Medium Scale Integration chips for a new version of the original P7 design, called Taurus. It was marketed in several versions within the generic name DPS-7 x5.

The P7G project was launched after the merger between CII and Honeywell-Bull in 1975 and had the main goal of being substituted to the high-end machine of CII, the Iris 80 systems (around 50 of them being then in the field). It also addressed the upgrade of the P7 systems (around one thousand were delivered in 1979). The new design of P7G was to support up to 4 processors in a true SMP configuration, to handle separately the central processor and the I/O channel functions and to have a hardware structure to support the level 64 type of instructions as well as the Iris 80 (itself derived from the SDS Sigma7).
While P7 did not implement pagination, the Iris 80 did it and a redesign of the TLB had to be made. P7G implemented more functions in hardware than P7, but it was still based on a microprogrammed engine. The changes in hardware lead to an almost completely new microinstruction code and so the P7G was from a hardware point of view a completely new machine.
The microcode storage was also changed because PROMs chips were no more available for CML control circuits. So the CPU firmware, stored in a fast static RAM CML store, had to be reloaded at system (and CPU) reinitialization by the service processor.
P7G featured also the first (in this product line) cache design. The cache was implemented in SRAM CML as an efficient store-into cache (possibly a first in the industry for a SMP system).

Trouble shooting of the processor by means of an oscilloscope was no more possible with micro packaging and a special mode of CPU operation was to interconnect all chips in a shift register being read and written (to reset the CPU) by the service processor. The clock frequency was set initially at 8 MHz.

The Leo system was delivered as DPS7-80 and DPS7-82 (the latter contained two processors). A DPS7-82 could be partitioned as two independent GCOS systems. That facility (probably a premiere in the industry) allowed a smooth conversion of Siris applications to the native mode. Although the CML technology dissipates more heat than TTL, Bull let its machines air-cooled (at the difference of IBM and Phoenix that use water-cooling for similar technologies).

the DPS-7-80 system

The Lyra system was a variation of P7G. Lyra allowed extending to 4 the number of processors and it was packaged in low cabinets sharing the appearance same design as the Taurus system.

In 1982, serious problems were encountered in the reliability of the micro packs. The price of gold (used for chips interconnection in P7G prototype) hiked significantly at that time and, to keep manufacturing costs down, it was decided to change the material of wiring to copper. Copper substrates showed short cuts and breaks due to the electro-migration of copper atoms in the ceramic material and the product line was saved by a come back to gold (the price of gold went down happily).

The peak park of the DPS-7 line was reached around 1985 with around 3000 systems, with a majority of them being Taurus systems. The cost of the CML technology remained high and micro packaging was no more applicable to a full range of computer products. After some hesitations (see hereunder) it was decided to reorient the DPS-7 into a CMOS technology.

The first product in CMOS was the Ares project introduced in 1987. The design started in the early 1980s as Archer a medium range computer implemented in NMOS. It was based on the P7G architecture adapted to the new technology. Eventually, the product was finalized in CMOS and the code name replaced by Ares. The CPU of the system was still multichip but it was the first Bull single-board mainframe computer, produced around one year ahead of IBM's similar product.

Ares was designed as a 6 ways SMP (shared milti-processor). The efficiency of multiprocessing in GCOS7 was very high, especially in a transaction environment that became at that time the bread and butter of mainframes systems.
The reliability of CMOS was considerably higher than previous technologies. That reliability made somewhat redundant the efforts made in RAS (reliability, availability, and serviceability) to support large configurations of DPS-7 due to the small number of parts and the low power consumption and related heat. Ares was introduced as DPS-7000 with the same software release as P7G that helps its introduction in 1987.

the DPS-7000 (Ares)

Ares was modified in 1988 to support paging as well as segmentation (Ares phase 2). A new entry version with a new peripheral architecture was also made available in 1988 (Ares phase 3 or Libra). TCheaper peripherals were coming from the microcomputer world (the most significant being SCSI discs). Instead of reimplementing the peripherals and the channel controllers in a proprietary way, the functions of the original architecture were performed by microprocessor based processor interfacing with the Ares I/O bus. A part from a shift to fixed sectors discs, the new I/O architecture was transparent to software.

Ares was not providing a reasonable upgrade path for the Lyra customers and a NEC central system (code named Aquila) was imported from NEC and delivered to customers in 1987. Its interior decor required the implementation of paging and GCOS7 software was modified to support it. Bull Ares and Lyra peripherals, sometimes different from NEC had to be connected to Aquila and checked with the new version (v3B) of GCOS 7. NEC had also to westernize the safety features of its CPU and to adapt to the western specific power supplies.

Aquila and Ares allowed a slight increase of the park of customers, but Aqula's cost and its transfer price (including the yen appreciation) did not provide a satisfactory upgrade for Ares customers.

So a new CPU design was started targeting an extension to 8 of the number of single board processors, taking advantage of the higher integration available. The Auriga design includes a deeper pipeline, wider data paths, and additional transfer from microprograms interpretation into hardware. A level2 cache was added shared by processors and I/O channels. The Auriga system was introduced as a high end DPS-7000.777

Auriga 2 a single-chip CPU

The same Auriga design was used in Auriga 2 (also known as Artemis) that features the whole CPU in a single chip of 7 million transistor, with a single board of 4 processors and their Level 2 cache. The initial shipping of Auriga 2 was a cost reduction of Auriga. Under the Artemis code name, the system was extended to up to 24 SMP processors. This set of processors can be partitioned for two independent systems and/or to specialize some sets of processors like Oracle database servers. The hardware architecture is still true SMP, but the specialization of some processors in dedicated servers (such as Oracle database server) or transaction processor improves the caches hit ratio and the overall performances.

In 1995, it was decide to enhance the DPS-7 product line with the capabilities offered in the Open Systems world. A system code named Diane was implemented. Diane was a system grouping Auriga 2 and Intel Pentium processors used as co-processors. That allowed to unload servers functions running open software in IA32 architecture (running UNIX or Windows 2000) and keeping Mainframe applications under GCOS7. This system allowed a flexible partitioning pioneered in P7G to distribute hardware resources between the cooperating operating systems.

A subsequent phase was to take profit of the performance improvements within Intel architecture to develop a DPS-7000 architecture emulator in IA32 and IA64. (Diane 2)

Paper designs

In addition, Bull creates several paper designs of processors of this product line.

The original design of the product line was the Charlie project in 1967. It concentrates exclusively on the architecture.
The R-530 project followed in 1968 as part of an overall General Electric APL advanced product line. An attempt was made to map the processor on standard blocks that might be VLSIfied in the future.Technology envisoned was initiallly CML, but it reverted to TTL when the project was renamed GE-730 in 1969.

In 1973-1974, after the introduction of Level-64, Honeywell-Bull and NEC worked to make an ECL version of P7, called P7-A. Honeywell canceled the project that was continued by NEC on their own.

The P7B project was the predecessor of P7G in 1975. Bipolar gate arrays were then considered. When the CII-Honeywell Bull merger became effective in 1977, the project was enhanced to the more powerful P7G using the micro packaging technology designed in GE semiconductor laboratory later transferred to Honeywell.

In 1982, Archer, a NMOS project called Archer was started in Les Clayes as a low-end system. Eventually, after the recognition of the CMOS superiority, the project evolved in the successful Ares project.

Bull embarks in 1983 (with Digital equipment) backing Gene Amdahl's Trilogy Corp attempt to use the wafer scale integration technology with the intent to use WSI for a high end system called Aquila. After Trilogy's failure, the project was cancelled and the code name was transferred to a non-related NEC-Bull project.

In 1987, a first attempt to develop a single chip DPS-7 processor called Altair failed to succeed.

NEC ACOS-4

The involvement of NEC in this product line has been already mentioned. NEC was a licensee of Honeywell since the 60s and has produced the H-200 product line Naturally, when Honeywell embarks in Series60, NEC followed and contributed by sending engineers to the project. In 1973, they signed a license agreement for P7, GCOS64 and the architecture with the right to build freely from that basis. The original Honeywell-Bull machine was introduced in Japan as ACOS-450 and the GCOS operating system was delivered as the first release of ACOS4.

NEC has made ACOS-4 systems the top of their mainframes competing with the most powerful IBM 3080 and 3090. This strategy was pursued even after the acquisition of the Toshiba business that included the license of GCOS8 36-bits systems that, in Bull and Honeywell, were the most powerful systems.

The privileged technology in NEC from 1979 to 1989 was the CML micropackaging, for which NEC developed several versions of custom-LSI.
NEC started in 1986 to build ACOS-4 systems in CMOS and had considered once to import Bull Auriga2 processor, but eventually they continued their own way.

Aquila system sold by Bull in 1986-1988 used the NEC S/750 processor.

It might be also worthwhile to note that NEC SX-series supercomputers used an ACOS-4 scalar processor (in addition to the main vector specialized processors).

NEC ACOS S/750

It could be noted also that NEC converted its ACOS-2 (alias Level-62, DPS-4) to ACOS-4 by an integrated emulator, while Honeywell and Bull transitioned them by software tools to GCOS-7 and UNIX.

NEC enhanced the original architecture first in introducing 2 and 4 ways multiprocessing, then by introducing the paging. Those enhancements were based on preliminary designs made with Bull in 1974.
Then in 1987-1988 a major extension called XSA (extended system architecture) was designed with Bull to overcome the limits of the original 32-bits architecture. If ACOS-4 was enhanced to support at least part of XSA, Bull did not invest enough money in GCOS-7 to take advantage of those improvements.

software I/O

Index