At the opening ceremony of GTC 2015, NVIDIA CEO Huang officially released the GTX Titan X graphics card and for professional market Quadro M6000. Last and this year’s graphics cards are based on 28nm Maxwell architecture, at the GTC conference, NVIDIA has announced a new roadmap for 2016 in which they will launch the next generation of Pascal architecture that supports 3D and NVLink memory bus, memory capacity will be upgraded from the current 12GB to 32GB and Pascal bandwidth will be 3X more than Maxwell.

NVIDIA Pascal architecture will be 10X faster than Maxwell

NVIDIA-Pascal

NVIDIA GPU Roadmap

Actually, in the last GTC conference they announced Pascal architecture with support of 3D memory and NVLink bus, so much new stuff on the GPU route, but this time NVIDIA announced more details on it, such as Pascal architecture SGEMM (FP32) to enhance the performance per watt compared to 66 percent Maxwell, now is an increase of 100% over last year.

Computing capabilities, NVIDIA architecture began to emphasize on Pascal FP16 and FP32 of hybrid computing, Pascal architecture upgrade tripled compared to Maxwell architecture at this point.

NVIDIA-Pascal-vs-Maxwell

NVIDIA Pascal 10X Faster than Maxwell

Pascal architecture will support 3D video memory, which is similar to HBM memory, but the latter is a 2.5D stack, so the risk is lower. NVIDIA also said that in Pascal architecture the memory capacity with reached 2.7 times the current level, currently GTX Titan X equipped with 12GB, 2.7 times is 32GB, which should catch up with the memory of the memory capacity, and then do not need to exchange data with the memory the bar.

In addition to a larger memory capacity than the memory bandwidth also need a revolution, NVIDIA launched NVLink bus in which bandwidth is tripled, while the GTX Titan X bandwidth is 336GB / s, then Pascal architecture will have three times the memory bandwidth in TB / s level.

NVIDIA® NVLink™ is a high-bandwidth, energy-efficient interconnect that enables ultra-fast communication between the CPU and GPU, and between GPUs. The technology allows data sharing at rates 5 to 12 times faster than the traditional PCIe Gen3 interconnect, resulting in dramatic speed-ups in application performance and creating a new breed of high-density, flexible servers for accelerated computing.

NVIDIA has also said that the above calculations are very rough, so take them with a pinch of salt.

Pascal architecture is expected to be launched in 2016, the manufacturing process will be upgraded to a FinFET, will be launched under the next generation of Volta architecture, the details are unknown.

NVIDIA-NVLINK-GPU-Scalability