Today at Nvidia’s GPU Technology Conference, Jen-Hsun Huang has announced Fermi, the company’s next generation GPU architecture.
Nvidia has long intended that their GPUs will evolve to look more akin to a CPU, a step at a time. Fermi certainly goes in this direction, leaving behind many of the more charming, but baroque, elements of GPU architecture. These and other changes demonstrate that Fermi is clearly aimed at the HPC market, which is a relatively new one for Nvidia.
The investment in programmability come at a cost though - the engineering resources, die area and power focused on HPC generally do not benefit the volume driver for GPUs - gaming. Nvidia must carefully treading a middle road - pushing GPU programmability for HPC opens up new markets and is
essential for the eventual convergence of CPUs and GPUs, but they must not stray too far from focusing on their core markets which provide most of the company’s volume and revenue. Whether they have fallen off this path will only be known when products are released into the market, and compared with the competition. However, it’s clear that this particular step carries significant risk with it, as many of the new features do not directly benefit graphics.