NVIDIA Reveals Details On Pascal GPU With Up To 16GB Of HBM2, 1TB/Sec Bandwidth

Although NVIDIA revealed some high level information on its next generation Pascal GPU architecture earlier this year, the company revealed some more interesting details this week at NVIDIA’s GPU Technology Conference in Japan. NVIDIA announced that Pascal will have close to a 2x performance-per-watt improvement over the current Maxwell GPU architecture, which is built on a 28nm process.

It’s also been revealed that Pascal will be produced using TSMC’s 16nm FinFET process, which will put it right up against AMD which is moving to use 14nm/16nm with two new Graphics Core Next (GCN) products in 2016. Other improvements including replacing PLX PCIe Gen3 bridging with NVLink, which enables bi-directional communications between two GPUs at 80GB/sec, up from 16GB/sec in NVIDIA’s current generation Maxwell GPUs.

And you didn’t really think that NVIDIA would let AMD hog all the spotlight with High Bandwidth Memory (HBM), did you? AMD’s HBM implementation features a much wider bus width than GDDR5 (1024-bits versus 32-bit) and employs vertical stacking to enable bandwidth of greater than 100GB/sec per stack. AMD’s Fury cards pack in 4GB of HBM with an effective memory bandwidth of 512GB/sec.

Pascal, however, includes High Bandwidth Memory 2 (HBM2), which will significantly increase available memory bandwidth compared to current generation offerings. Flagship consumer Pascal boards will ship with four 4GB modules, combining to provide 16GB of memory with maximum bandwidth of 1TB/sec. However, it doesn’t end there — professional level cards, like the Quadro and Tesla series, will be able to support up to 32GB of memory while touting internal bandwidth of an incredible 2TB/sec.

The first Pascal boards are expected to ship during the opening half of 2016.