Nvidia unveils Pascal, its next-gen GPU

Pascal a super computer module was unveiled by Jen Hsun Huang, CEO of Nvidia, at the annual conference of the company. This new architecture optimizes data transfer between the CPU and the GPU and uses a new type of memory arranged in three dimensions. According to NVidia, the module presented, which is one third the size of a PCIe graphics card brings the power of a supercomputer in the area occupied by two credit cards.

 NVidia, the manufacturer of graphics processors for computers and systems on a chip (SoC) that is found in tablets and Smartphones, held its annual conference this week (CWG 2014) in San Jose, California. This is an opportunity to unveil new graphics processor architecture for the next generation of chips, video games, mobile SoC devices and supercomputers. Pascal called in honor of the French scientist Blaise Pascal, this super computer is an important technical milestone with the introduction of several technologies that optimize the use of RAM and reduce bottlenecks for data transfer between the GPU and the central processor CPU.

The first of these is the new 3D memory, which consists of stacked DRAM cells arranged vertically and integrate all on the same semiconductor as the CPU. NVidia announces that the memory should be multiplied by 2.5 with a much larger bandwidth. This 3D architecture also allows you to place voltage regulators near the chip, which has the effect of optimizing the energy efficiency that would be four times higher than a GDDR5 memory solution. NVidia emphasizes that through this, the Pascal module will have one third of the size of a PCIe graphics card.

For its new processor architecture, NVIDIA opted for a RAM in 3D. The memory access is much faster while allowing a more compact design. Another key innovation is a new interconnected technology between the CPU and GPU named NVLink. It was co-developed with IBM with the goal to leverage the speed of data transfer between the two processors. According to NVidia , NVLink is 5-12 times faster than that can offer PCI Express bus third generation, because it leaves the GPU memory access the CPU at full bandwidth and vice versa.

The founder announced a bandwidth of 80 to 200 Gb / s, against 16 GB / s of PCIe 3.0. NVLink also renews the interconnection between the GPUs themselves, which will make creating clustering configurations that may include multiple GPUs. This new type of interconnects will greatly benefit applications that take advantage of hardware acceleration and massively for parallel computing. Through this system of unified memory, NVidia promises developers they will be released for optimizing problematic communication protocols.

IBM plans to integrate this technology into future versions of Power processors. Note that NVidia does not intend to abandon the PCIe bus. All transactions and configuration control initiated by the CPU go through a PCIe connection only. If a CPU is not NVLink supported, may interconnect between the GPU and function to improve the performance rise, or scaling. In other words, NVLink will somehow technical manufacturers of processors may or may not choose to adopt. The first chips from the architecture Pascal arrive in 2016.