Cray T3E

Cray T3E. The Cray T3E was a pioneering massively parallel supercomputer series developed by Cray Research in the mid-1990s. It was a direct evolution of the earlier Cray T3D, designed to deliver significantly higher performance and scalability for scientific computing and engineering simulation. The system became a dominant platform in high-performance computing throughout the late 1990s, powering major research at institutions like the National Science Foundation and the United States Department of Energy.

Overview

The T3E project was initiated to address the growing demand for parallel processing capabilities beyond what the Cray T3D could provide. It was formally announced in late 1995, with the first systems shipped in 1996 to leading research facilities. The design philosophy centered on a scalable distributed memory architecture using a high-performance interconnection network. Key figures in its development included Steve Chen and teams at Cray Research's Eagan, Minnesota facilities. The system competed directly with other contemporary parallel systems from IBM, SGI, and Intel.

Architecture

The core architectural innovation was its three-dimensional torus interconnect, which provided low-latency communication between thousands of processing nodes. Each node contained a DEC Alpha 21164 microprocessor, a hallmark of the design, paired with custom Cray logic for network and memory management. The memory system was a distributed shared memory architecture, facilitated by hardware support for barrier synchronization and atomic operations. This design eliminated major bottlenecks present in the Cray T3D, such as memory bank conflict and network contention, enabling more efficient execution of parallel algorithms.

Models and specifications

The product line evolved through several models, starting with the initial T3E featuring a 300 MHz DEC Alpha processor. The T3E-600 increased clock speeds to 600 MHz, while the T3E-900 and T3E-1200 represented peak performance iterations. System scale ranged from modest 8-processor configurations to the largest installation, a 1,088-processor machine at the Pittsburgh Supercomputing Center. The flagship T3E-1200, deployed at sites like the National Energy Research Scientific Computing Center, achieved a peak performance of over 1 teraFLOPS. Each node's local DRAM capacity grew from 64 MB to 2 GB across the series.

Software and programming

The primary operating system was UNICOS/mk, a distributed version of Cray's UNICOS derived from Berkeley Software Distribution kernels. Programming models were centered on message passing interface and SHMEM library for explicit parallel coding. The system also supported High Performance Fortran and Co-Array Fortran for data-parallel applications. Key software tools included the Cray CF90 and Cray C++ compilers, along with the TotalView debugger. This environment enabled major codes in computational fluid dynamics, climate modeling, and computational chemistry to achieve unprecedented scalability.

Impact and legacy

The T3E series had a profound impact on the adoption of massively parallel processing within the scientific community. It powered groundbreaking research in astrophysics, materials science, and genomics at laboratories like Lawrence Livermore National Laboratory and the Max Planck Society. Its commercial success solidified Cray's transition from vector processor to parallel architectures. The technological lessons directly influenced subsequent Cray designs, including the Cray X1 and the Cray XT series. Several decommissioned T3E systems are preserved in collections like the Computer History Museum, symbolizing a pivotal era in supercomputing history.

Category:Supercomputers Category:Cray computers Category:Massively parallel computers Category:1995 in computing