Scientific Supercomputing for Performance

This area is the proving ground and the driver for innovative architectural techniques.

Microprocessors have made tremendous gains in floating-point performance
- high clock rates,
- pipelined floating point units (e.g. multiply-add every cycle)
- instruction-level parallelism
- effective use of caches (e.g. automatic blocking)
The process is under way for large-scale multiprocessors to replace vector supercomputers.

Let us take a look at uniprocessor performance versus parallel performance on the LINPACK package.

Uniprocessor Performance: of supercomputers and microprocessor-based systems on the LINPACK benchmark. Performance in MFLOPS for a single processor on solving dense linear equations is shown in the following figure for the leading CRAY vector supercomputer and the fastest workstations on a 100 x 100 and 1000 x 1000 matrix. (ref: Parallel Computer Architecture, Culler et al.)
Parallel Performance: of supercomputers and MPPs on the LINPACK peak performance benchmark. Peak performance in GFLOPS for solving dense linear equations is shown in the following figure for the leading CRAY multiprocessor vector supercomputers and the fastest massively parallel processors (MPP) systems. (ref: Parallel Computer Architecture, Culler et al.)

The above graph shows that even vector CRAYs became parallel: X-MP (2 to 4 proc), Y-MP (8 proc), C-90 (16 proc), T94 (32 proc). Also CRAY produces MPPs such as (T3D and T3E).