Apparently the new Pascal architecture from NVIDIA will bring twice as many transistors as current Maxwell architecture found on Titan X.
As the successor of today’s main architecture used on NVIDIA’s high-end graphics cards found on GM200, the new design was suspected to increase twofold. The huge increase seems to come from Pascal’s 16nm FinFET process and its transistor size that will be two times smaller.
Being manufactured at TSMC, NVIDIA’s GPUs share the same process manufacturing with AMD’s graphics cards. Since TSMC has announced that it will start moving towards 10nm process manufacturing in 2016, we can expect major performance increases in two years from now. Although it’s unlikely NVIDIA will wait until 10nm processes enter mass production in 2017, it will still count on their 14nm process manufacturing that will come with FinFETs pretty soon.
The FinFET branding depends on which side of the transistor you look at. Although the gate is identical on both process manufacturing techniques, the fin source is tighter and flows straight through the gate offering a fully depleted operation, even though the gate size remains the same for both 16nm and 14nm process.
In addition to that, NVIDIA will use second-generation HBM memory stacking technology from SK Hynix with approximately 32GB on the top-tier card. The HBM2 will allow 8Gb per DRAM die, 2Gbps speed per pin and 256 GB per second Bandwidth/ stack. By comparison, the first-generation HBM found on AMD’s latest Radeon R9 Fury offers 2Gb Density per DRAM die, 1Gbps speed per pin, 128 GB/s GB per second Bandwidth and maximum of 4 Hi stack chips with 4GB per HBM card.
HBM2 will allow cards manufacturers to bump the memory size per HBM2.0 die from 1GB per chip to four, or even eight, result in massive memory sizes stacked on the GPU die of 16GB to 32GB respectively. Pascal will probably bring both versions on its next-generation GP100 GPU.
Fudzilla believes the new GPU will come as early as 2016. Since NVIDIA announced a month ago that it had already taped out its Pascal architecture GPUs, we can rest assured that no immediate delays will push back the advent of Pascal.