Latest player on the parallel supercomputing scene is New Bedford, Massachusetts start-up International Parallel Machines Inc, which last week showed off its revolutionary IP-1 computer in Boston. Key features of the machine are claimed to be that all processors can access a shared memory simulataneously without arbitration, regardless of the number of processors, and the price-performance: the company says that the IP-1 is less than $1,500 per MFLOPS in raw performance, and under $4,000 per MFLOPS on the 64-bit Linpack benchmark. And it says that in terms of scalar performance, a $180,000 sys tem can be used as a back-end to a minicomputer, in which role it can sort 1m numbers in one second. Other features of the IP-1 include Flash Data Transfer, an interconn ection network that enables a proc essor to pass 1Mb of data to anoth er processor in under one microsec ond; a database with a 32-bit addr ess range which since the machine uses 64-bit words can extend to 4G-words – that’s 32Gb, and can be extended to 48-bit addressing, where we are up into the Teraword addressing. The company says that over 200 vector, matrix and paral lel routines are also included in the standard library for scientific and engineering applications. Ext ernal communications include serial port, Ethernet or optional bus-to bus interface at up to 80Mbytes-per-second. The thing supports optional D-Scan GR4416 terminals for three-dimensional multi-user colour graphics comparable with single-user super workstations. No word yet on the operating soft ware or when the company plans to ship the thing – or what it is using as building block processor.
This article is from the CBROnline archive: some formatting and images may not be present.
CBR Online legacy content.