Article 631 of comp.sys.cdc: Newsgroups: comp.sys.cdc Path: mri-gw!psinntp!psinntp!psinntp!rutgers!att-out!pacbell.com!amdahl!netcomsv!netcom.com!smryan From: smryan@netcom.com (S M Ryan) Subject: Re: Help with Cyber Architecture Message-ID: Organization: Santa Clara Research Park X-Newsreader: TIN [version 1.2 PL1] References: <1994Sep8.133911.32913@waikato.ac.nz> Date: Thu, 8 Sep 1994 22:49:07 GMT Lines: 92 No Cyber 207, unless that was the Cyber 2XX which was spun off into GF10 of ETA Systems. It started as the Star100, which I've heard was an internal competitor with Cray's Cyber 8600 before he left. It gave birth to Star100D. About this time CDC renamed all its machines CYBERs, so it gave birth to Cyber203, and then the 205. The 205 successor, the 2XX, was spun off to ETA. The series used 64-bit words, twos complement, ASCII. 64 bit single precision floats, double precision similar to the 6600. 48-bit integers. 64-bit addresses were divided into 16-bit length field and a 48-bit virtual BIT address. Real arithmetic was not IEEE. 16 bit exponent, unbiased I believe, and 48-bit fraction. A machine zero was represented by an exponent 8XXX. An indefinite value had an exponent of 7XXX. The machine was designed to use 'significant' arithemetic. The result of and addition/subtraction was not normalised with the leading zeros being significant. Multiplication preserved leading zeros in the product. Customers demanded normalised arithmetic. A single precision add/subtract normalised was used in Fortran, and Peili Wu eventually got normalised double precision working. A 32-bit halfprecision was also available. Integer arithmetic only included 48-bit add/subtract. Real multiply/divide were used, similar to the 6600 method. The scalar unit had 256 64-bit registers. The first 128 registers could also be addressed as 256 32-bit registers, the upper and lower halves of the 64-bit registers. Register 0 was always a zero. The first 32 registers were assigned an interface convention. No data caches. Up to 3 addresses were sent to the memory interface which eventually sent back the value. Bit, byte, halfword, and fullword access. The scalar unit was divided into various functional units. Except the divider, all functional units were pipelined. The vector unit was divided into pipelined functional units with their own memory interface. Vector registers were not used, these were memory-to-memory vector units. Eight independent memory banks were simultaneously addressed reading eight words at a time and piping them through. Similarily for the output. The maximum vector length was 65535 elements. Vectors could be 64-bit integers, 64-bit or 32-bit reals, or bit-vectors. A bit vector could be used to turn off and on a vector operation on a per element basis. Up through the 203, bit-vectors could be used as extra long integers, allowing, for example, 65535 bit division. These machines also allowed BCD numbers of up to 65535 digits. Although these would be useful today for national debt computation, at the time they were superfluous and dropped in the 205. (The Star dated back to when NCR and CDC were working together. The Star was to be a scientific and business supercomputer.) The vector unit also allowed various reshapings of vectors, such as compression/expansion, gather/scatter. The only the competitor was the Cray machines which used vector registers and short (64 element) vectors. This difference let the Crays start up a vector operator much faster, but harder to maintain long vectors. On applications that used short vectors and lots of scalar, Cray was faster. Applications that used mostly long vectors were faster on 205s. On Cray machines, vectors as short as three elements were faster as vector operations than scalar operations. On 205s, because of the additional overhead of starting and stopping memory access, the break even point was closer to 20 to 60 elements. The operating system was not Unix (or even close). It came from Lawerence Laboratories, radition burns and mutations intact. The languages were assembly, fortran 66 extended/77 extended, and a system implementation language halfway between fortran and assembly. Purdue later wrote a C compiler. For the 2XX, CDC was building a new compiler technologies for that and the 180s. When that project contracted to just the 180s, 2XX people left and made their own company to provide ETA with compilers. My best to Peili Wu, Darlene Ha, Steve Maziar, Tom Woo, Azar Hashemi, Jing Yang, Dave Torgerson, Toni Roberts, Dave Tweten, and Bob Ng. (Those were the days, my friend/we thought they'd never end.) -- What is the cardinality of | smryan@netcom.com PO Box 1563 the set of the cardinals? | Cupertino, California ... ond lof-gearnost. | (xxx)xxx-xxxx 95015