For Sameer Halepete, the challenges of staying ahead of Moore’s Law are only making chip design more fun. “In the past, you had ideas to improve power, cost, performance…people would just say, go to the next process node. Now that the option of going to the next node is further and further out, the value of those innovations and ideas that get us the performance, power, and area improvements has grown,” he told an audience of Design Automation Conference attendees during his Tuesday, June 7, keynote. Halepete, VP of VSLI Engineering at NVIDIA, titled his talk, “Driving the Next Decade of Innovations in Computing.” And what’s driving the next decade is fun indeed. We’re talking applications like virtual reality, deep learning, and self-driving cars—applications that can dramatically change the way we live, work, and play. NVIDIA has been refining its graphics processors for more than 20 years. The work they’ve done for gaming applications has yielded results that are being applied to an array of other compute-intensive applications, such as building design and space exploration. During his talk, Halepete discussed some key applications for NVIDIA processors. Virtual Reality The ability to simulate reality has improved immensely over the years. Halepete showed a rendering of recently retired NBA player Kobe Bryant from an NBA Live 96 video game. Bryant was essentially a pixelated stick figure, and the image had no depth of field and very little definition. Next to it was another rendering of Bryant, this one from the 2016 version of the game. Huge difference: the image of Bryant was much like a high-resolution photograph. Visual computing has evolved such that “we can make you feel like you are actually there in the scene,” he said, noting that NVIDIA offers a comprehensive set of GPUs, graphics drivers, and a software development kit for virtual reality designs. Deep Neural Networks Over the last couple of decades, engineers have honed the ability to do parallel computations and get high compute density on a die. These are also the ingredients to accelerate scientific research. Deep learning applications, in particular, have benefited from the advancements. Neural networks running on GPUs can now be trained to perform pattern recognition. Speech recognition, long an area of difficulty for computing, has improved such that we’re now seeing living-room devices that people can interact with. Oak Ridge National Laboratory uses NVIDIA GPUs in its powerwall display , which supports scientific visualization projects. Self-Driving Cars Every year, there are roughly 1.2 million traffic-related fatalities around the world. Self-driving cars, said Halepete, can potentially eliminate the vast majority of these tragedies. NVIDIA’s automotive team tested an autonomous vehicle equipped with a deep learning neural network over 3,000 miles. Initially, the car couldn’t avoid road obstacles, but after 3,000 miles, it learned. Said Halepete, “It was not taught what a road is, it was not taught what other cars look like. It was given the visual input and it was told what a normal driver would do.” Managing Design Complexity Designs certainly continue to grow more complex to provide the level of functionality available. Yet, Halepete noted that although NVIDIA has increased the number of transistors it has taped out by a factor of 10 since 2010, it has done so without having to substantially increase the size of its team. Fine-tuning of implementation techniques and applying lots of optimizations makes a difference. So, too, does collaboration within teams and with EDA partners who develop tools to accelerate design and verification time and timing and power optimization. Your customers, he told the DAC audience, don’t care about the slowing of Moore’s Law. They care about meeting their performance, power, and area targets. “Each team needs to understand what the other team truly cares about,” he said of the importance of collaboration. As further encouragement, Halepete noted that the semiconductor industry isn’t the only industry to face a challenge such as Moore’s Law. At one point, commercial aircraft were doubling their number of passenger miles every four years for 16 years, until the late 2000s brought a dramatic slowdown. Yet, Boeing continued to innovate, experimenting with different materials and composites and improving the inside of its cabins. From 1970 to today, Boeing’s value has increased 100-fold, Halepete noted. ”I’m convinced there are many such ideas available to us,” said Halepete. “I’m looking forward to working with many of you to make our many-fold increase happen with us.” Christine Young
↧