During the week of SC13, NVIDIA launched its highest-performance graphics processor yet, the Tesla K40 GPU accelerator. Compared to the chipmaker’s previous high-end Kepler, the K20X, the Tesla K40 sports more memory, higher clock rates, and a greater number of CUDA cores. Now that a couple month have passed since the debut, we’ll examine the impact of these improvements in the case of three real-world applications.
Yesterday, NVIDIA rep Roy Kim blogged about three demanding applications that have experienced gains thanks to the newest Keplers. They are weather forecasting, social media feed analysis and financial risk modeling. “These burly applications gobble up computing power like an NFL linebacker plowing through a box of Ding Dongs,” as Kim says, so it’ll be interesting to see how they’ve benefitted from the K40 GPU.
Detailed weather forecasting presents a particularly thorny computational challenge. To predict weather conditions across the country, the U.S. National Weather Service relies on an application called WRF, which it feeds to powerful supercomputers. University of Wisconsin-Madison researcher Dr. Bormin Huang has been using the new K40 GPUs to accelerate WRF in order to generate more accurate, higher resolution forecasts in less time. Based on his findings, the Tesla K40 runs the application nearly 30 percent faster, compared with the K20 – and with the K40’s new GPU Boost feature enabled, that figures shoots to 46 percent.
The high-end graphics processor is also being used to analyze and extract useful data from popular social media feeds, such as Twitter. A Massachusetts startup, Map-D, tracks tweets and provides real-time analysis of trends. Despite the massive volume of data involved, the database application is hosted on a single server outfitted with eight Tesla K40 graphics processors.
From the blog:
With an aggregate GPU memory of 96GB, it’s large enough to store 1 billion tweets in memory for ultra-fast analysis. And with GPUs, it’s over 70 times faster than with CPUs alone, or fast enough to let you know what’s happening in the Twitterverse at any given moment in time.
With enough in-memory storage to hold one billion tweets, the interactive application can provide insight into any trending topics in real-time even as the data changes from moment to moment. From the Sochi Olympics to the Oscars or even the next #POTUS, Map-D can map it.
The next highlighted GPU-accelerated HPC application takes us to the world of finance, where Monte Carlo algorithms are used to evaluate complex financial instruments, minimizing the risk associated with such transactions. Software vendor Xcelerit has long relied on hardware accelerators, including GPUs, to speedup compute-intensive applications for the financial services community. The Xcelerit team tested the K40 and the K20X using the Monte-Carlo LIBOR swaption portfolio pricer as the yardstick, and found that the Tesla K40 is consistently 20-30 percent faster.
If you want to see how your application responds, NVIDIA offers a free test drive.