ssors and expands to a maximum of 128. RAM configurations range from 128 MB to 32 GB. I/O support includes fast-and-wide SCSI-2 and FDDI.
The basic building block of the Exemplar is a hypernode that holds a maximum of eight processors. Each hypernode is essentially an SMP system where the processors and memory connect through a special four-by-five, high-speed, nonblocking crossbar bus. Gallium arsenide gate arrays implement the crossbar logic. When a processor accesses RAM, bus logic called an
agent
(see the figure
"The Convex Hypernode"
) looks for an idle bus to memory and makes the connection. The fifth part of the crossbar is for I/O only.
Convex achieves scalable parallel processing by tying hypernodes together. Four one-way toroidal rings connect the hypernodes. Called Coherent Toroidal Interconnect (CTI) by Convex, it's based on the IEEE standard 1596-1992 Scalable Coherency Interface (SCI). The crossbar logic monitors
memory accesses, and local (i.e., on-board) access remains on the hypernode. Remote (i.e., off-board) accesses get passed to the rings. Through CTI, the memories on all the hypernodes appear as one globally shared memory space (see the figure
"Toroidal Interconnect"
).
Because the rings are actually cables, you can tie several Convex systems together to obtain more processing power. Furthermore, you still have one unified memory space, even though the RAM is physically located in different machines.
The Exemplar uses a multiprocessor verion of Unix called SPP-UX. SPP-UX is binary-compatible with HP-UX, so the Exemplar can draw on thousands of HP workstation applications.
A basic Exemplar 1200/CD 2 with two processors and 128 MB of RAM costs $128,320. A basic Exemplar 1200/XA 8 has eight processors and 256 MB of RAM and costs $349,500. Each four-processor upgrade will set you back $107,800.
illustration_link (8 Kbytes)

A crossbar minimizes bus contention for RAM.
illustration_link (17 Kbytes)

The CTI organizes distributed memory into one shared space.