Nvidia’s GPU technology Conference took place in the USA recently, with a new flagship GPU (see Product News) announced as well as a brand new architecture.
Maxwell is Nvidia’s 10th-generation GPU architecture, and will be succeeded by ‘Pascal’ in 2016. According to Nvidia, Pascal is as much as 10 times faster than Maxwell.
Pascal’s design was developed thanks to Nvidia’s work on deep (aka machine) learning. GPUs based on the architecture will have three key features: mixed-precision computing; 3D memory; and NVLink.
Mixed-precision computing enables Pascal-based GPUs to compute at 16-bit floating point accuracy; this is of particular benefit to classification and convolution, which are key to deep learning.
Memory bandwidth limits the speed at which data can be delivered to the GPU. 3D memory will increase bandwidth by three times, with almost three times the frame buffer capacity of Maxwell-based GPUs. Pascal-based GPUs will have their memory chips staked on top of each other and placed adjacent to the GPU, rather than further down the processor boards. Bits will thus need to travel a small distance between the two components, accelerating communication and raising power efficiency.
Finally, Nvidia’s NVLink interconnect enables faster data movement: between 5 and 12 times faster, the company says, than PCIe. It also means that the number of GPUs in a system working together in deep learning computations can be doubled. Nvidia has a video about NVLink that can be viewed at http://tinyurl.com/q48pn9n.
Deep learning is a key part of the automotive market. We talked to Nvidia about its move into this space at CES, where it was showing the new Drive PX platform (Nvidia’s Teraflop Chip Enables Self-Aware Cars). Drive PX, which now has a price ($1,000), monitors a car’s surroundings to enable self-driving vehicles. It uses two Tegra X1 processors and can combine data streams from up to 12 cameras.
Automotive is the fastest-growing segment of Nvidia’s Tegra business, with higher gross margins than the devices segment. Revenues almost doubled YoY in Q4’14, and more than 7.5 million cars now use Nvidia’s technology – well up from 4.7 million a year ago.
At the GTC, Tesla CEO Elon Musk predicted that self-driving cars would become the rule, rather than the exception, within 20 years.
Interestingly, Anandtech reports that Nvidia has significant advantages over AMD in the GPU market. The site’s most demanding benchmarks (Crysis 3 at UltraHD in high quality, with FXAA) saw the Titan X beat the R9 290X’s performance by 54%; AMD’s high-end GPUs had their last significant refresh in 2013. For many years, the two company’s have been almost even in performance. (TA)