NVIDIA has fully announced their Pascal GPU architecture at GTC Japan, confirming that it will offer 2x the performance per watt of their previous gen efficient Maxwell architecture and will contain a max of 16GB of HBM2 memory for consumer GPUs.
Nvidia’s Pascal GPUs will be built using TSMC’s 16nm process, which is not just a smaller processing node, but a process which uses FinFET transistors, which will give a great performance in gaming,leap over the 28nm planar transistors used in current generation GPUs.
The move to 3D memory from conventional planar GDDR5 memory will not only give a huge leap in GPU memory performance, but it will also allow increased capacities in a great scale, with 4x4GB HBM2 memory giving consumer Maxwell GPUs up to 16GB of VRAM.
NVIDIA’s Marc Hamilton had this to say about HBM and Nvidia’s NVlink:
“Using 3D memory, not only the memory capacity will go up, the memory bandwidth will go up significantly. With a much faster GPU, and higher memory bandwidth, the existing interconnects in the server are just plain outdated. So, we had to develop our own interconnect called NVLink, five times faster than existing technology.”
The next generation GPUs from both AMD and Nvidia will be very interesting really, with both sides saying that it will give a 2x jump in performance per watt, shifting to FinFET transistors and shifting to use HBM 2.0 memory.
These next generation GPUs are expected to come in market from begining to mid 2016