Emerging technologies like artificial intelligence (AI), machine learning (ML) are driving more than just the semiconductor market. They also drove the conversations I had with everyone I spoke with at and around SEMICON West 2018. Turns out people had lots to talk about.
AI and Machine Learning
I often think back to the 2010 MEMS Executive Summit in Scottsdale Arizona, when Karen Lightman declared, “sensors are fricken’ everywhere”, and talked about how cool it would be when our smartphone learned all our habits and became less a phone and more a personal assistant. Here we are, eight years later, and not only has that scenario become a reality, it’s become our expectation.
In his SEMICON West keynote, Mark Papermaster, AMD, talked about how we now demand better experiences that require more compute power. We want better interactions that anticipate what we need. We want the experience to be immersive and surround us. That will happen thanks to augmented reality, virtual reality, and middle reality, he said.
“Machine learning is a new era promising opportunity for us all in the industry,” declared Papermaster. “Semiconductor devices that accelerate machine learning algorithms are now at cost levels that can enable mass adoption.”
Good news for 3D
Thomas Urhmann, EV Group, says AI, ML and blockchain are definitely impacting the 3D side of EVG’s business. These applications are driving new CPU architectures, such as Google’s tensor processing unit (TPU), in order to process data from the cloud faster. Because of high data rates and latency issues, companies are choosing 3D for implementation. “Before it was a cost game, now it’s performance driven,” he said. “If it has the benefit, it doesn’t matter what the cost is.”
EV Group’s big announcement at SEMICON West was its achievement of 50nm alignment accuracy in its Gemini FB XT, which is a big deal, as the previous generation tool accuracy was 200nm. Target applications include memory stacking, die partitioning, and backside illuminated (BSI) image sensors.
This is important particularly for image sensors as the data rate continues to rise, and smarter devices are needed. Integrating BSI with memory allows for faster cache rate, global shutter, and better integration at the pixel level.
It’s not the Data, it’s what you DO with it
Leti’s new CEO, Emmanuel Sabonnadiére, explained it best. It’s not the volume, but the quality of the data that matters. He explained the logic behind Leti’s neuromorphic approach to AI for edge computing.
“There’s a ridiculous amount…an explosion of data and nobody seems to care,” he said. Rather, we need to be more “brain-inspired” – in our memory, we only keep the data that makes sense.
“It’s good to forget things,” he explained. “We store unnecessary data. Data will fuel the industry but finding the right data for the right decisions is what’s needed.” I couldn’t agree more!
Hubert Lakner, Fraunhofer Microelektronik agrees with Sabonnadiére about the push to edge computing. Take, for example, autonomous vehicles, that need to react immediately, and can’t wait for the information to go to the cloud and come back to the device. Instead, he envisions an environment comprising Lidar and image sensor systems that require AI and machine learning at the edge to process the data and make decisions in real time. His colleague, Juergen Wolf, Fraunhofer IZM-ASSID, said this requires 3D and high-density interconnects, and the reliability and thermal aspects need to be included. All of this points towards the need for heterogeneous integration technologies.
Because, as Wolf puts it, “Heterogeneous integration is more complex than just stacking layers,” the Fraunhofer Mikroeletronik and Fraunhofer IZM helped to establish the Research Fab Mikroelectroniks Deutschland (FMD), a consortium of 13 European (mostly German) R&D centers focused on the development of silicon-based technology, compound semiconductors, and heterogeneous integration (Figure 1). The consortia presented its own keynotes in its booth during SEMICON West.
While imec seems to have set 3D aside to chase 3nm and EUV lithography, Leti and Fraunhofer continue to pursue 3D architectures and systems integration.
12nm is Good Enough
According to Sabonnadiere, Leti has made a strategic decision to support technology down to 12nm, and then innovate on the substrate to leverage 3D architectures, because 12nm is “good enough”. Fully-depleted silicon-on-insulator (FD SOI) technology, which enables both 3D and heterogeneous integration technology, is being explored as a substrate for Leti’s 3D monolithic approach known as “CoolCube” because its temperature is manageable, integration works well, and as a 3D application, its never been explored by the industry, he said.
Sabonnadiere also talked about Leti’s recent five-year agreement with Soitec to pursue fully depleted silicon-on-insulator (FDSOI) R&D. Soitec’s success in SOI is ramping up, and the company is committing facility to full production to meet the demand of the growing FDSOI market, he said. All of their R&D work, therefore, will move to Leti. Mark Lepedus, SemiEngineering, published an in-depth interview with Sabonnadiére here.
That’s it for me, from SEMICON West 2018. If you missed part one of SEMICON West conversations, its right here. ~ FvT
