Cray T3D

from Wikipedia, the free encyclopedia

The T3D (abbreviation for Torus , 3-dimensional ) was Cray's first attempt in the field of massively parallel supercomputer architecture. It was introduced in 1993 and was also Cray's first supercomputer to be based on a non-proprietary microprocessor architecture.

The T3D consisted of 32 (up to 2048) Processing Elements (PEs) , each of which included a DEC Alpha 21064 (EV4) microprocessor clocked at 150 MHz and RAM, consisting of 16 or 64 MB DRAM , a. The PEs were grouped in pairs in so-called nodes and connected with a 6-way switch. These switches provided a maximum bandwidth of 300 MB / second in both directions and were connected to one another in such a way that the topology of this network corresponded to a three-dimensional toroidal body .

The T3D was developed in such a way that a Cray Y-MP model E, M90 or C90 acted as a "front-end" system or host and all I / O and other system services were processed via its UNICOS operating system . The PEs themselves ran with a simpler UNICOS / MAX microkernel .

The T3D was available in different configurations:

  • The SC models ("SC" for "Single Cabinet", only one housing was used) shared the housing with the Y-MP host computer and could be equipped with 128 or 256 PEs.
  • The MC models ("MC" for "Multiple Cabinet", several cases could be used) accommodated the T3D in one or more cases, which were liquid-cooled . The host computer was installed in its own housing.
  • In contrast to this, the MCA models ("MCA" for "Multiple Cabinet, Air cooled") were air-cooled and took between 32 and 128 PEs.
  • There was also an MCN model, which also had liquid cooling, but was wired differently than the other models and thus also allowed an uneven number of PEs.

The T3D was replaced in 1995 by the faster and more advanced T3E .

  • T3D MC 256 at EPFL

  • T3D MC 256

  • Opened CRAY-T3D at EPFL

  • Front view of the closed CRAY-T3D at EPFL

  • Control panel of a T3D MC 256

  • Detail of the control panel

  • Look inside