Nvidia is planning to release two more general-purpose GPUs over the next three years, as the graphics chipmaker increasingly boosts the performance of its products for the high-performance computing market.
Nvidia Chief Executive Jen-Hsun Huang says the two GP GPUs, codenamed Kepler and Maxwell, would debut in 2011 and 2013, respectively. Huang made the announcement Tuesday during his keynote at the company's GPU Technology Conference in San Jose, Calif., according to a live blog of the keynote on the Nvidia's website.
Huang also announced the release of Fermi, the latest generation of Nvidia's compute unified device architecture. CUDA allows programmers to use the C programming language to code algorithms for execution on the GPU. In high-performance computing, that part of an application that's heavily computational would be compiled using Nvidia's CUDA compiler in order to run on a server or workstation. Nvidia claims CUDA is not difficult for developers because it uses C, a familiar programming language.
The Fermi GP GPU features up to 512 CUDA cores, compared to the 240 cores in its predecessor Tesla. In addition, Fermi has error-correcting code that's expected to improve Nvidia GPUs for use in data centers. ECC protects data-sensitive applications, from medical imaging to financial options pricing, from memory errors.
Kepler, based on a 28-nanometer manufacturing process, is expected to offer three to four times the performance per watt of Fermi. Nvidia has pledged to spend $2 billion in developing the next-generation GP GPU.
Maxwell will be an advancement over Kepler in that it will offer more capabilities, such as virtual memory. In addition, Maxwell will perform at a level 16 times greater than Fermi, or 16 gigaflops per watt, Huang said.
Nvidis is not alone in adapting graphics technology to the HPC market, which includes the medical, financial and oil and gas industries, as well as scientific research. Advanced Micro Devices offers GP GPUs through its ATI graphics unit.
The general-purpose graphics chips are fairly new to the industry with the vendors introducing the technology roughly four years ago. How well GP GPUs will do against other chipmakers, such as Intel and IBM, remains to be seen.
Nevertheless, adoption of Nvidia's technology appears to be growing. Many computer makers in the HPC market have announced systems based on the company's GP GPUs, including Appro, Bull, Cray, Dell, Hewlett-Packard, IBM, SGI, Supermicro, T-Platforms and Tyan.