Nvidia shows pascal specs

Nvidia shows Pascal's specs. As it has been doing with each new architecture, Nvidia has finally released its official document from Pascal in which it reveals numerous characteristics of its imminent and promising graphic architecture.

Pascal specifications, an important advance against Maxwell

The most powerful Pascal- based GPU will be the GP100 which is made up of a total of 60 SM units that form the basic structure of Pascal and that will be cut as other lower-range chips are manufactured. This GPU will debut with the Tesla K80 professional card.

Each of the SM drives contains a total of 64 FP32 single precision CUDA cores, a clear difference from Maxwell and Kepler which contain 128 and 192 cores respectively, in the case of a full Gp100 GPU we have 3, 840 single precision cores. In addition, each of these SMs is divided into two blocks consisting of 32 nuclei each, which are joined by an instruction buffer, a task scheduler and two execution units, characteristics shared with Maxwell.

We now focus on Pacal's double precision calculation capabilities and find that each SM unit has 32 FP64 double precision cores. A full GP100 GPU has a total of 1, 920 dual -precision CUDA FP64 cores so the ratio between FP32 and FP64 is 2: 1.

We now turn to Pascal's L2 cache and it turns out that the GP100 core has a total of 4, 096 Kb of this extremely high speed memory, a great advance compared to the 1, 536 Kb of the GK110 or the 3, 072 Kb of the GM200. With this increase in the L2 cache, it is possible to reduce the number of accesses to VRAM memory, which translates into less need for bandwidth, less consumption and, what really matters, higher performance.

If you want to see more details about Pascal's specifications you can access the official document here

Source: videocardz