Supercomputing in a Shoebox: The Convergence of High-Performance Computing (HPC) and High-Performance Embedded Computing (HPEC)
Moderator: Jeremy Kepner, Massachusetts Institute of Technology, Lincoln Laboratory
Panelists: Jack Dongarra, Professor, University of Tennessee; Charles Holland, Director of Information Technology, DUSD (S&T); Craig Lund, CTO, Mercury Computers, Inc.; Charles Seitz, CEO & CTO, Myricom, Inc.; Tony Skjellum, President, MPI Software Technology, Inc.; Tom Sterling, California Institute of Technology
The majority of the world's computing is performed in embedded systems (wireless apps, network routers, set-top boxes, etc.). Larger and larger parallel embedded systems are being deployed and the world of High-Performance Embedded Computing (HPEC) is increasingly dealing with traditional HPC issues (bisection bandwidth, shared vs. distributed memory, scalable software, etc.). The increasing physical size of today's largest HPC systems is problematic for all but a few users. Increasingly the HPC community is dealing with traditional HPEC issues (Mflops/ft3, Mflops/watt, etc.). This panel will introduce these issues and will discuss how they relate to both the HPC and HPEC communities. Questions that will be addressed include:
DENSITY: HPEC systems provide a 10x increase in Mflops/ft3 over their HPC counterparts. Is the need for compute density limited to special-purpose applications, or will high-density computing be required for next-generation supercomputers to be feasible?
CONVERGENCE: HPC systems are increasingly becoming more focused on density while HPEC systems are becoming more focused on programmability. Will these two regimes converge?
GROWTH: Both HPC and HPEC are driven by commercial and DoD markets (servers, medical, wireless, etc.). Will one of these areas become so dominant as to create a "victor"?