William J Dally, chief scientist and senior vice president of Research at Nvidia Corp, delivered a keynote speech at the 46th Design Automation Conference, which is taking place from July 27, 2009, in San Francisco, the US.
The theme of the speech was "The End of Denial Architecture and the Rise of Throughput Computing." And he said that architectures making most of parallelism and the locality of data like GPUs will become effective while CPU performances are reaching a ceiling.
For example, the GTX280, Nvidia's existing GPU, is mounted with 240 floating point arithmetic cores (stream processors), realizing as high a performance as 1TFLOPS with one chip. Dally predicts that GPU performance will continue to grow by 70% per year due to the continuing miniaturization.
"In 2015, the 11nm-generation will have 5,000 floating point arithmetic cores, a performance of 20TFLOPS per chip with an operating frequency of 3GHz, and a memory bandwidth of 1.2 Tbytes per second," he said.
GPUs are used not only for graphics but also for financial calculations, fluid analyses, medical picture processing and other applications. Nvidia is planning to expand the use of GPUs in the fields of scientific computation and embedded devices too.
Also, in the field of EDA, especially in applications such as timing verification and correction and verification of mask patterns (e.g. OPC and DRC), the performance demand is growing faster than that indicated by Moore's law, making it necessary to ulitize GPUs, Dally said.
From the viewpoint of EDA tool users, he cited the estimate of power consumption in architecture exploration as well as design support to make most of the locality of data and traces as issues concerning GPU design and requests to the EDA industry.