Nvidia Corp. has announced "Carma" development kit to developers, who want to design software that relies on ARM general-purpose processors and Nvidia graphics chips compatible with CUDA architecture. With Carma, developers will be able to try designing software for energy-efficient high-performance computing projects. Unfortunately, Carma does not support double precision computing and thus will hardly be adopted by serious developers.
Nvidia Carma - CUDA on ARM - is based on Tegra 3 system-on-chip with five general-purpose ARM Cortex-A9 processing cores and various multimedia functionality with 2GB of DDR3 memory, as well as Nvidia Quadro 1000M (GF108GL) graphics processing unit with an array of 96 CUDA [compute unified device architecture] stream processors and 2GB DDR3 memory. According to Nvidia, Carma offers 270GFLOPS single-precision performance. The development kit is based on Linux Ubuntu Derivative OS.
Earlier this year Nvidia and Barcelona Supercomputing Center (BSC) said that they had initiated development a new hybrid supercomputer that, for the first time, will use energy-efficient, low-power Nvidia Tegra ARM system-on-chips as well as high-performance Nvidia Tesla compute accelerators. The building blocks for the supercomputer will be similar to Carma development kit.
The Carma development kit is aimed at those, who would like to develop supercomputing applications for energy-efficient exascale systems, which, in Nvidia's point of view, will be based on ARM general-purpose cores and accelerated using Nvidia's highly-parallel compute engines. The kit itself does not look too impressive though as it features Fermi-generation GF108GL graphics chip that does not support double precision computing, a crucial technology for high-performance computing (HPC) applications. Even though Carma may be good for research, it will barely demonstrate actual performance advantages of CUDA on ARM technology.
Nvidia Carma development kits will be available starting from Q2 2012.