
04.30.24-Startup-Ecosystem-Conference-Startups-Sangtera

-
Interactive transcript
TAIRAN WANG: Good afternoon. I'm Tairan Wang. I'm co-founder and CEO of Sangtera. Sangtera is a spinoff from MIT Lincoln Laboratory. We are building a technology to help build AI chips faster. So over the last two years, we have all witnessed the incredible rise of generative AI and, with it, a rise in the demand for more computing power.
This means that, in order for all these applications to work, they need a different kind of chips. What makes these chips uniquely suited for AI is a new 3D architecture that placed a large amount of memory in close proximity to the logic engines in vertical stacks. Well, put chips vertically means that they can be much closer to each other. The shorter interconnect leads to reduced latency, lower power consumption, and tighter integration. Increased I/O counts means increased bandwidth, not to mention the high yield and IP reuse that's associated with the chiplet approach in general. And that's exactly what the 3D AI chips from Nvidia and their competitors are trying to deliver to their customers.
Now, here's a problem. They can't build these chips fast enough because 3D chips require a new set of manufacturing steps, including the precision stacking of chips. To achieve that position with existing equipment means that they need to run a lot slower, lowering their output from 20,000 chips per hour to less than 2,000. And that's because the precision-stacking heads are too slow and too bulky to fit more than two at a time as illustrated in this video.
This bottleneck of stacking chips for AI is the biggest challenge on the horizon for the industry. It is one we are uniquely equipped to address. Our technology allows many chips to be attached and aligned in parallel, greatly increasing the throughput to more than 20,000 per hour matching that of wafer processing tools.
And the key to this massive parallelism is a new precision tiny stage, packing kilograms of alignment machinery into the volume smaller than a penny. And inside each stage is a new kind of electric actuator that's powered by the surface tension of water. You may remember the surface tension. That's a liquid interface property that allows, for example, insects to walk on water.
It's one of the strongest force in the microscale. Now, we have figured out how to harness this force, packing thousands of engineered droplets into a small volume delivering forces that's hundreds of times that of conventional actuators of the same size. And this is the key innovation that makes the tiny precision stages possible.
So generative AI requires a large amount of compute, which relies on the new 3D architecture. And that needs precision packing of chips in volume. And that is a productivity challenge that Sangtera can help to solve.
The foundation of this technology has been published in leading journals and is also patented. We have, so far, demonstrated high-force density motor, high-precision stepping, and the ability to move chiplet and other payloads. We are currently working to demonstrate additional key performance metrics with support from an NSF STTR grant and also a fellowship from E14.
We are looking for additional partners in advanced material and fabrication to help improve this technology and also to scale it for production. We are also looking for partners in semiconductor advanced packaging to learn more about their needs and product requirements to take this revolutionary capability into the market to drive productivity for the making of next generation AI processors. Thank you.
-
Interactive transcript
TAIRAN WANG: Good afternoon. I'm Tairan Wang. I'm co-founder and CEO of Sangtera. Sangtera is a spinoff from MIT Lincoln Laboratory. We are building a technology to help build AI chips faster. So over the last two years, we have all witnessed the incredible rise of generative AI and, with it, a rise in the demand for more computing power.
This means that, in order for all these applications to work, they need a different kind of chips. What makes these chips uniquely suited for AI is a new 3D architecture that placed a large amount of memory in close proximity to the logic engines in vertical stacks. Well, put chips vertically means that they can be much closer to each other. The shorter interconnect leads to reduced latency, lower power consumption, and tighter integration. Increased I/O counts means increased bandwidth, not to mention the high yield and IP reuse that's associated with the chiplet approach in general. And that's exactly what the 3D AI chips from Nvidia and their competitors are trying to deliver to their customers.
Now, here's a problem. They can't build these chips fast enough because 3D chips require a new set of manufacturing steps, including the precision stacking of chips. To achieve that position with existing equipment means that they need to run a lot slower, lowering their output from 20,000 chips per hour to less than 2,000. And that's because the precision-stacking heads are too slow and too bulky to fit more than two at a time as illustrated in this video.
This bottleneck of stacking chips for AI is the biggest challenge on the horizon for the industry. It is one we are uniquely equipped to address. Our technology allows many chips to be attached and aligned in parallel, greatly increasing the throughput to more than 20,000 per hour matching that of wafer processing tools.
And the key to this massive parallelism is a new precision tiny stage, packing kilograms of alignment machinery into the volume smaller than a penny. And inside each stage is a new kind of electric actuator that's powered by the surface tension of water. You may remember the surface tension. That's a liquid interface property that allows, for example, insects to walk on water.
It's one of the strongest force in the microscale. Now, we have figured out how to harness this force, packing thousands of engineered droplets into a small volume delivering forces that's hundreds of times that of conventional actuators of the same size. And this is the key innovation that makes the tiny precision stages possible.
So generative AI requires a large amount of compute, which relies on the new 3D architecture. And that needs precision packing of chips in volume. And that is a productivity challenge that Sangtera can help to solve.
The foundation of this technology has been published in leading journals and is also patented. We have, so far, demonstrated high-force density motor, high-precision stepping, and the ability to move chiplet and other payloads. We are currently working to demonstrate additional key performance metrics with support from an NSF STTR grant and also a fellowship from E14.
We are looking for additional partners in advanced material and fabrication to help improve this technology and also to scale it for production. We are also looking for partners in semiconductor advanced packaging to learn more about their needs and product requirements to take this revolutionary capability into the market to drive productivity for the making of next generation AI processors. Thank you.