Pascal was first announced as a mystery GPU between Maxwell and Volta at last year’s GTC. Tuesday’s announcement gives us the first concrete details of Pascal.
Huang said that Pascal, which is set to arrive in 2016, would have a ten-fold overall average improvement over Maxwell, and a four times boost in mixed-precision workloads. As far as performance per watt, it will offer a two-fold performance over Maxwell. However he later cautioned this was “CEO Math” and actual performance may vary.
Pascal will use a suite of new technologies, including 3D-stacked memory and NVLink (Huang says it will offer a five-fold improvement over PCI-E). It will be built on TSMC (TPE: 2330) 16nm FF+ (FinFet plus — the follow-up to FinFET) process node. It will also use High Bandwidth Memory, allowing a three-fold improvement of bandwidth for its 32 GB of RAM.
Cards with Pascal will likely be marketed towards CUDA workstations, or perhaps as some sort of competitor to Intel’s (NASDAQ: INTC) Xeon Phi co-processors. Game developers have a tough time pushing the limits of current generation cards as it is.
Pricing and availability of Pascal-based cards will be available later this year or early next year.