Fujitsu VP2000
From Wikipedia, the free encyclopedia
The VP2000 were the second series of vector processor supercomputers from Fujitsu. They replaced their earlier Fujitsu VP series machines starting in 1990, and sold for a short period before being replaced by various massively parallel machines.
The VP2000 was similar in many ways to their earlier designs, and in turn to the Cray-1, using a register-based vector processor for performance. For additional performance the vector units supported a special multiply-and-add instruction that could retire two results per clock cycle. This instruction "chain" is particularly common in many supercomputer applications, allowing the VP2000 to outperform the Cray-1 at a similar clock speed.
Another difference is that the main scalar units of the processor ran at half the speed of the vector unit. According to Amdahl's Law computers tend to run at the speed of their slowest unit, and in this case unless the program spent most of its time in the vector units, the slower scalar performance would make it 1/2 the performance of a Cray-1 at the same speed. The reason for this seemingly odd "feature" is unclear.
One of the major complaints about the earlier VP series was their limited memory bandwidth -- while the machines themselves had excellent performance in the processors, they were often starved for data. For the VP2000 series this was addressed by adding a second load/store unit to the scalar units, doubling memory bandwidth.
Several versions of the machines were sold at different price points. The "low end" VP2100 ran at an 8ns cycle time and delivered only 0.GFLOPS (about 4-8 times the speed of a Cray), while the VP2200 and VP2400 increased this to 4ns and delivered between 1.25 and 2.5GFLOPS peak. The "high end" VP2600 ran at 3.2ns and delivered 5GFLOPS. All of the models came in the /10 versions with a single scalar processor, or the /20 with a second, while the 2200 and 2400 also came in a /40 configuration with four. Due to the additional load/store units, adding additional scalar units improved performance by increasing memory bandwidth, as well as allowing several programs to run at the same time and thereby increase the chance there was something to process on the vector unit. Each unit is said to increase performance 1.5x, allowing the VP2400/40 to match the performance of the earlier VP2600/20.
The machines were supplied with either the Unix-compatible UXP/M or the VMS-compatible VSP/S, both supplied by Amdahl. The later was used for Fortran programs while the former was typically used for C, and vectorizing compilers were supplied for both languages.
Like most companies, Fujitsu turned to massive parallelism for future machines, and the VP2000 family were not on the market for very long. Nevertheless over 100 were sold.
