Neural network simulations on massively parallel computers: Applications in chemical physics

Conference · Fri Jan 01 04:00:00 EST 1993

OSTI ID:6649368

Sumpter, B G; Getino, C; Noid, D W ^[1]; Guenther, R E ^[2]; Halloy, C ^[3]

---

1. Oak Ridge National Lab., TN (United States)
2. Nebraska Univ., Omaha, NE (United States). Dept. of Physics
3. Tennessee Univ., Knoxville, TN (United States). Joint Inst. of Computational Sciences

A fully connected feedforward neural network is simulated on a number of parallel computers (MasPar-1, Connection Machine CM5, Intel iPSC-2 and iPSC-860) and the performance is compared to that obtained on sequential vector computers (Cray YMP, Cray C90, IIBM-3090) and to a scaler workstation (MM RISC-6000). Peak performances of up to 342 million connections per second (MCPS) could be obtained on the Cray C90 using a single processor while the optimum performance obtained on the parallel computers was 90 MCPS using 4096 processors. Efficiency such as these has enabled neural network computations to be carried out for a number of chemical physics problems. Several examples are discussed: multi-dimensional function/surface fitting, coordinate transformations, and predictions of physical properties from chemical structure.

• Research Organization: Oak Ridge National Lab., TN (United States)
• Sponsoring Organization: DOE; USDOE, Washington, DC (United States)
• DOE Contract Number: AC05-84OR21400
• OSTI ID: 6649368
• Report Number(s): CONF-9306101-1; ON: DE93006739
• Country of Publication: United States
• Language: English

Related Subjects

664000* -- Atomic & Molecular Physics-- (1992-)
74 ATOMIC AND MOLECULAR PHYSICS
99 GENERAL AND MISCELLANEOUS
990200 -- Mathematics & Computers
ARRAY PROCESSORS
BENCHMARKS
CHEMICAL PHYSICS
NEURAL NETWORKS
PARALLEL PROCESSING
PERFORMANCE
PHYSICS
PROGRAMMING