Nvidia announces Turing architecture for GPUs

Nvidia has announced the Turing architecture for GPUs on Siggraph. In addition, the company has announced some graphics cards, but they are not intended for the consumer market.

In its presentation on Siggraph, Nvidia emphasizes ray tracing, the technique in which, for example, a large amount of light rays are ‘shot’ from the position of the camera in different directions. How these light rays are then reflected is calculated on the basis of the color and properties of the material. Turing GPUs will have specialized RT cores on board to enable ray tracing with minimal delay.

Together with a new generation of Tensor cores optimized for neural networks, a Turing GPU should render a physical world six times faster than the previous Pascal generation.

The new Turing GPUs can use gddr6 memory at a maximum of 14Gbit/s. That is what Samsung, SK Hynix and Micron all offer, although Samsung now also offers faster memory. Given the number of transistors of the GPUs just announced, the process is probably the same as with the Quadro GV100, 12nm-FinFet from TSMC, which was released this spring.

Nvidia announced four different models, the Quadro RTX 8000, 6000, 5000 and RTX Server. The RTX8000 has 4608 cuda cores and 576 Tensor cores on a die size of 754 square millimeters. The GPUs are for the professional market and the RTX 8000 will cost ten thousand dollars. The new GPUs will be released this fall.

Nvidia is expected to present the consumer versions of the Turing GPUs at the GamesCom game fair in Cologne next week. They will follow the current GTX10xx series. It is unknown when exactly they will be released.