Recently, the risc-v instruction set architecture has received a lot of attention. This new open source instruction set allows processor developers to easily develop various types of chips. NVIDIA joined the risi-v foundation long ago and did a lot of research. NVIDIA researchers recently presented a paper at the 2019 VLSI circuit symposium that USES the risc-v instruction set to develop a multi-chip modular extensible deep neural network accelerator, and published an abstract on their website.

Picture from NVIDIA

NVIDIA says deep neural networks require high performance, accuracy and power consumption, and that building a deep neural network accelerator is often difficult and expensive. So they used low-power, high-bandwidth interconnect technology to connect a single inference accelerator chip with all sorts of computing power. NVIDIA 16 through integration in a chip chip internal network connection for the Processing of deep learning operation element (Processing Elements, PE) and a controller adopts RISC instruction set - V, and a single chip can provide highest 4.01 TOPS (trillions of operations per second), and NVIDIA researchers through the network connection between the highest 36 chip, provide up to 128 TOPS of work force.

Picture from NVIDIA

In addition to NVIDIA's published chip calculations, its researchers also provide information such as the size of the chip. Using TSMC's 16nm process, the cumulative core area of a single chip is 3.1 square millimeters, while the cumulative Die area is 6 square millimeters, and the core power consumption is 0.03w to 4W. The cumulative core area of 36 chips (6×6 specifications) is 111.6 square millimeters, the cumulative area of Die is 216 square millimeters, and the core power consumption is between 5W and 100W (a little large span).

Picture from NVIDIA

Even with its high power consumption, NVIDIA's 36 TOPS per square millimeter, according to the NVIDIA paper. Unfortunately, NVIDIA isn't planning to release this chip, just to show off its high-efficiency chip design, which NVIDIA may incorporate into its products in the future.

News from：超能网