Fermi’s 16 streaming multiprocessors (SM) are positioned around a common L2 cache. Each SM is a vertical rectangular strip that contains an orange portion (scheduler and dispatch), a green portion (execution units), and light blue portions (register file and L1 cache).

GPGPUs Invade Supercomputing 
The Future at Work
 

In the past decade, standard x86 processors went from near-zero to hero, largely replacing RISC processors to populate 70 percent of high performance computing (HPC) systems. Starting in 2010, IDC predicts that general purpose graphics processing units (GPGPUs) will make headway. Multiple 2010 procurements include GPGPUs alongside x86 processors, and Oak Ridge National Laboratory has announced plans for a mega-supercomputer based on NVIDIA’s “Fermi” GPGPU co-processors. GPGPUs also are growing their footprints in industries such as oil and gas, financial services and bio-life sciences.

Make no mistake: x86 processors will dominate for the foreseeable future, but GPGPUs are being deployed for development work, and the transition to production computing will inevitably follow. Despite improvements, the x86 per-core bandwidth (bytes/flops ratio) keeps declining as the number of cores per processor multiples. Tuned-down x86 core speeds also throttle performance. Meanwhile, GPGPUs have become more advanced and easier to program. For the right applications, GPGPUs promise to provide more bang for the buck. Today, GPGPUs are typically co-processors linked to x86 or other standard base processors. In the future, they may all be equals on the same die.

Steve Conway is the Research Vice President in IDC’s High Performance Computing group. He may be contacted at editor@ScientificComputing.com.