Cray Research Inc’s new massively parallel Alpha RISC-based T3D machines duly appeared yesterday in a variety of sizes. At the small end is a 32-node version rated at 4.8 GigaFLOPS peak, and costing $2.2m, this skips up through 64-, 128-, 256- and 512-processor versions to a box with 1,024 Alphas that is rated at 153.6GFLOPS and costs $31m. At the very top of the line is an un-priced 2,048 processor machine offering 307.2 GFLOPS peak. Key to the machines’ speed, says Cray, is the inter-processor latency, which has been whittled down to less than one microsecond. As reported (CI No 2,262) the new architecture is designed to keep Digital Equipment Corp on its toes by making it relatively easy for Cray to switch processor suppliers if the Alpha does not keep coming up with the goods – Sparc is ruled out currently since it lacks a 64-bit implementation. The machines run Unicos Max, a distributed, parallel version of the existing Unicos vector operating system, which looks like Unix to the operator. Each of the processors runs the Unicos microkernel, written in C, which is also one of the languages available to application programmers. The more likely option is Cray’s extension to Fortran 77 which supports both traditional message-passing and data parallel programming styles. The T3Ds implement the first stage of a three phase process; the next step is due in two or three years with a machine dubbed the T3E which, Cray claims, will offer peak performance in the TeraFLOPS range, the third phase, following a couple of years after that is aiming at sustained TeraFLOPS speeds. The company has started a joint venture called Aptos with Lyon, France-based partner, Jacques Stern’s Advanced Computer Research International SA, ACRI (CI No 2,156). Aptos will act as a high-tech dating agency, bringing together user firms with software partners to solve particular applications problems and implement them for the Alpha-based machines. Nine orders have already been signed for the machines, and Cray says it is involved in between 20 to 30 invitations to tender at the moment. Cray reckons that in five years time around 50% of its customers could have left the C90 vector processing architecture in favour of the massively parallel approach – the speed of the migration depends in large part of the speed with which the proper programming algorithms can be designed to duplicate 20 years of experience at vector programming. At the same time the company renewed its commitment to produce a next-generation vector machine codenamed Triton around mid-decade.
