GPU architecture not suited for AI, says Xilinx CEO
Increasingly automated vehicles and traffic systems, smart cities and the complex manufacturing landscapes of the future all have one thing in common: They have to be able to deal with an “explosion of data”, Peng said. And not just data: Unstructured, heterogeneous data that have to be processed under stringent real-time requirements. Today’s standard CPUs are less and less suitable for dealing with this data storm. “They simply are not good at processing unstructured data,” Peng judged.
But Peng’s verdict does not only hit the plain vanilla processors that power today’s PCs, laptops and embedded computers. The rising use of Artificial Intelligence in cars, robots, and in all kind of real-time systems requires new computing approaches as well. Today, it can be observed that more or less every automotive OEM and all major tier ones have established strategic relationships with vendors of AI computing platforms, which in terms of architecture typically are based on Graphics Processing Units (GPUs). Due to their high parallelism, GPUs are regarded as an excellent choice for AI-related tasks. This however, is only a small part of the picture, Peng said. The reason: While this parallelism makes GPUs ideally suited to implement neural networks and thus are a good candidate for the training part of the AI game, they are far less suited for the productive use of such applications, for inference applications. The reason: “GPUs are not good at latency”, Peng said. And latency – or better, the absence of latency – is a critical requirement in most safety-related real-time applications such as ADAS, automated driving or robotics.
But this is not the only downside of GPUs in cars and similar mobile real-time systems. Another one – and perhaps one that is more owed to real-world implementations than to the architecture as such – is its power consumption. “In series vehicles, you see much more stringent power requirements,” Peng said, referring to current popular lab-optimized computing platforms. Plus, to have success in the automotive business, semiconductor vendors need a zero-fault strategy and a long product life – qualities a newcomer chip vendor needs to acquire before being accepted in the demanding automotive industry.
Obviously, these remarks referred to competitor Nvidia which lately seems to be the darling of the carmakers and tier ones. But how about Xilinx’ own track record in the automotive business? Until recently at least, Xilinx was known as provider of prototype silicon – perhaps with interesting performance features, but too costly for series production?
This perception is a thing of the past, Peng responded. “Today, we are in series production vehicles. We ship 40 million units into automotive markets. This is all production (vehicles), not just prototypes”, Peng made clear.
Besides Nvidia, Intel is another major chip vendor who tries to get a foot into door of the automotive business. Actually, Intel made major strategic moves to increase its standing in this industry. For instance, it acquired vision processing company Mobileye and entered a strategic cooperation with BMW to develop a self-driving vehicle. So far however, Intel’s impact in the automotive market is negligible, at least from Peng’s perspective. “I did not notice much change in the market,” Peng said, adding “We noticed more impact from the GPU side.”
Which brings us back to the subject of Nvidia. “We see good chances to displace GPUs” Peng said. “Actually, we are already doing that.”
Continental relies on Xilinx expertise for auto-driving computer platform
Automotive-qualified MPSoC family targets ADAS platforms