Gerhard Klimeck and colleagues reach 1.44 petaflops on ORNL’s Jaguar

At left, schematic view of a nanowire transistor with an atomistic resolution of the semiconductor channel. At right, illustration of electron-phonon scattering in nanowire transistor. The current as function of position (horizontal) and energy (vertical) is plotted. Electrons (filled blue circle) lose energy by emitting phonons or crystal vibrations (green stars) as they move from the source to the drain of the transistor.

Researchers from Purdue University, the University of Alabama–Huntsville, and Switzerland’s ETH Zurich are finalists for this year’s coveted ACM Gordon Bell Prize, thanks to a nanoscale simulation of electronic devices performed on Oak Ridge National Laboratory’s Jaguar supercomputer.

Purdue’s Gerhard Klimeck and colleagues reached 1.44 quadrillion calculations a second, or 1.44 petaflops, using 221,400 of Jaguar’s 224,000 processing cores.

The $10,000 prize is administered by the Association of Computing Machinery and funded by high-performance computing (HPC) pioneer Gordon Bell. It goes each year to developers of the world’s most advanced scientific supercomputing application. Klimeck’s team is one of five finalists for the prize. Winners will be named November 17 in Seattle during the SC11 HPC conference.

Mathieu Luisier, a former research faculty member in Klimeck’s Purdue research group now at ETH, was the principal developer of OMEN. The application is a computer-aided design tool that uses the laws of nonequilibrium statistical quantum mechanics to simulate, at the scale of individual atoms, the behavior of transistors that are at the heart of microprocessors. This approach has become necessary because of the success chip manufacturers have enjoyed in shrinking the transistors on electronic devices. At some points these transistors are as narrow as 100 to 150 atoms.

As a result, models of these devices that rely on classical physics necessarily fail to encompass the devices accurately. OMEN captures quantum behaviors such as energy quantization, quantum mechanical tunneling, the wave nature of electrons and holes, and the atomistic granularity of the devices.

The team simulated two realistic two-dimensional nanoelectric devices: a single-gate high electron mobility transistor (HEMT) and a double-gate band-to-band tunneling field-effect transistor (TFET).

“I believe that HPC applications are dominated by science questions,” Klimeck said. “OMEN is the first electrical engineering software, maybe even the first engineering code at all, that runs at the petascale. It allows us to explore electronic design spaces with repeated simulations, rather than a few hero runs.”

With double-precision simulations—using 64-bit values—it reached a sustained 1.27 petaflops for HEMT and 1.28 petaflops for TFET. The team also performed a mixed-precision simulation of HEMT—using both 64-bit and 32-bit values—and reached 1.44 petaflops.

Team members credited on the Gordon Bell Prize paper are, besides Klimeck and Luisier, Timothy Boykin of the University of Alabama–Huntsville and Wolfgang Fichtner of ETH.