At the conference, SC15, dedicated to supercomputers and high-speed computing, NVIDIA has revealed more details about the generation of graphics solutions Pascal, thus confirming the preexisting rumors.
Processors embody Pascal in TSMC silicon is using the technology of 16 nm FinFET. According to the calculations of our foreign colleagues in the crystal Pascal will fit up to 17 billion transistors. GPU will use the memory HBM second generation with a capacity of 1 terabyte per second, initially its volume is 16 GB, but then NVIDIA will produce solution with 32 GB of memory on board. For multiprocessor configurations NVLINK provides an interface with a capacity of 5-12 times faster than PCI Express.
Pascal programming interfaces will support DirectX 12 and Vulcan. By the way, gaming solutions are unlikely to be completed with a lot of memory, and according to some information, limited to 4 GB volume. It is important that NVIDIA architecture Pascal seeks to cater not only to individual gamers, but the supercomputer industry. Thus, the potential for future GPU computing will grow significantly: Pascal processor can provide performance at the level of 4 teraflops with double precision, and is expected to 10 teraflops with single-precision numbers.
Interestingly, the generation processors Volta, following Pascal, according to the development plan should ensure that the performance of NVIDIA at 7 teraflops with double precision. Maxwell's successors will debut in the first half of next year, but Volta get to the HPC market only in 2017, while the gaming segment - in 2018. Related Products :
|