The name Pascal, for Blaise Pascal, the French mathematician and physicist (and more) who lived from 1623 to 1662, has been used for a programming language and now a high-performance graphics processing unit (GPU) architecture from NVIDIA. With models including up to 150-plus billion transistors fabricated at 16 nanometers, the Pascal-based GPUs have been eagerly (and some might say “long”) awaited by system builders and high-performance computing (HPC) experts alike. The first two consumer graphics cards, the GeForce GTX 1080 and GTX 1070, are now entering the channel.
Karel Filipe, founder and owner of Lotus Computer USA LLC of Altamonte Springs, Fla., is a big fan. “We’ve been building a lot of custom desktops with the new NVIDIA GeForce GTX 1080 and 1070. We’re also putting these GPUs in our laptops, with significant performance and efficiency improvements over the last-generation 9-series chips,” he says.
“Gamers and professionals who don’t opt for a [NVIDIA] Quadro GPU are seeing really impressive performance with these GPUs,” continues Filipe, “and the computers are running a bit cooler, thanks to the enhanced efficiency of the Pascal cards.” And cooler for a custom laptop is critical, since it’s the most difficult heat-control environment for a high-end graphics card. Filipe says the GTX 1070s he uses in laptops are desktop chips that are underclocked for cooling reasons.
HPC groups have been anxiously waiting for the GPUs as well. According to Karl Freund, senior analyst, HPC and deep learning (a branch of machine learning), at technology analyst and advisory firm Moor Insights & Strategy, “Pascal is NVIDIA’s next-generation platform, for which there are several implementations. All of them support 1/2 precision floating point (16 bit) natively, as well as 8-bit integer math. These operations are now two and four times faster than previous [Maxwell] implementations at the instruction-set level. Then you add more processor cores, the faster frequency, etc. So the speed-up is significant for deep-learning applications.”
Virtual Reality Performance
With regard to the GTX 1080, NVIDIA claims it offers twice the virtual reality performance of its previous top-end card with three times the energy efficiency. It also uses Micron GDDR5X memory, which is said to provide a 10 Gbps memory pipe. Moreover, NIVIDIA asserts that a 500-watt power supply will support a desktop PC with a single GTX 1080 card.
Clients in video production and gaming who have been begging for more performance will take little persuasion to upgrade. And one Pascal-based graphics card will be enough, according to Filipe. “I think the days of having three or four GPUs in one desktop are coming to an end. Even two GTX 1070s are overkill. NVIDIA has done a really impressive job and right now nothing compares.”
The other graphics hot-button is virtual reality. Filipe believes these cards will push client acceptance. “We just completed a pretty big order specifically for VR using a GTX 1080 in each custom PC we built. I expect a lot more demand for VR in the coming months.”
In the realm of deep-learning training, Freund recommends NVIDIA’s Tesla P100, the top end of the company’s GPUs that it calls “the most advanced data center GPU ever built.” These support two interfaces: PCIe and NVLINK, NVIDIA’s high-speed bidirectional interconnect. For use with NVIDIA’s deep-learning inference solution, Freund suggests the P4 and the P40. These are built for hyperscale HPC projects.
Freund also addresses the “long-awaited” tag associated with Pascal processors. “It typically takes two or three years to develop a new processor family, so I don’t see this as being so long as some say. It’s normal.” Of course, this is the technology biz, where miracles are supposed to ship this quarter, not next. “The next generation, Volta, will be out by the end of 2017,” he adds.
The GTX 1080 retails for $599, while the GTX 1070 retails for $380. Founders Edition cards retail for $699 and $449, respectively.