Currently, NVIDIA still relies on Pascal architecture for their GPUs, including for the GTX 10 Series. However, since a few years ago, they have announced the successor architecture of Pascal, called Volta. Some time ago, NVIDIA finally announced their new product based on Volta architecture, again not a GPU for desktop, but Tesla accelerator.
Use 12 nm & HBM2 Fabrication
The latest member of the Tesla family, with the Volta architecture, introduced by NVIDIA is the Tesla V100. In accordance with the code, this accelerator will continue the task of Tesla P100. The so-called Tesla V100 will offer a single-precision compute process capability of up to 15 TFLOPS, an increase of 10.6 TFLOPS offered by Tesla P100.
NVIDIA mentions that the Tesla V100 is manufactured with 12 nm FinFET fabrication from TSMC. Tesla V100 is equipped with 16 GB HBM2 which offers bandwidth of 900 GB / s. This accelerator carries 5120 CUDA Core and runs at a speed of 1455 MHz.
Rely on NVLink
NVLink is still relied on NVIDIA to support communication between Tesla V100 modules and systems. However, unlike NVLink on Tesla P100 that “only” offers bandwidth of 160 GB / s, NVLink on Tesla V100 can offer up to 300 GB / s bandwidth. This is, of course, important to support the acceleration process in the data center, which is the main target of Tesla V100 users, especially in the field of machine learning and deep learning.