Efinix, a semiconductor start-up based in Santa Clara, is trying to create a programmable chip that can be customized for machine learning used in applications like robotic arms, autonomous drones, security cameras, and driverless cars.
The goal is to make these chips small and efficient enough for factory sensors and even smartphones, which usually defer machine learning tasks to the cloud. Efinix is betting on a novel FPGA, which contains logic blocks connected through programmable interconnects to form digital circuits.
The firm was started by Sammy Cheung, a former senior director of Altera’s custom silicon unit and vice president at programmable chipmaker Stretch. The other founder was chief technology officer Tony Ngai, a chip architect who helped design commercial FPGAs for Xilinx, Lattice, and Altera, which Intel acquired for $16.7 billion in 2015.
Since its 2012 founding, the company has eschewed raising large amounts of venture capital to build its new architecture, called Quantum. Efinix almost shipped its first programmable chip two years ago but decided against it to further develop the technology. “This is not easy,” Cheung, the company’s chief executive, told Electronic Design.
“We didn’t think we would last that long. We are experienced enough to know that it takes a long time, so no need to rush it,” Cheung said, adding that Efinix would continue to benefit from the slowing of Moore’s Law, the chip industry’s instruction manual and the source of Xilinx and Altera’s long reign in programmable logic.
Traditionally, FPGAs are divided into separate routing and logic blocks arranged in a checkerboard pattern. The individual cells are interconnected in different ways depending on the application. These chips are typically bulky and expensive because it requires lots of extra circuitry to ensure that the logic blocks can be stitched together in the most complex way possible.
Efinix built a cell that can perform either logic or routing functions, allowing chips to be manufactured four times smaller and two times more efficient as traditional FPGAs. Cheung said that the company’s XLR cell would enable devices to run machine learning locally instead of using remote servers, where the algorithms are typically trained on GPUs.
“We can also build logic for whatever performance we need,” Cheung said. Efinix plans to release chips targeting autonomous cars down to industrial sensors in the second half of 2018. Further out, the 20-person firm plans to sell server chips that handle the inferencing phase of machine learning, in which algorithms apply what they have learned.
Efinix claims to cut down on interconnect resistance that has worsened as the transistors inside FPGAs continue to shrink. To limit resistance, chipmakers are putting between 10 and 14 metal layers inside advanced chips, while Efinix manufacture chips with only seven slivers of metal. That means lower costs and easier integration with other chips like application-specific integrated circuits (ASICs).
These edits could make programmable chips more widespread. Traditional FPGAs are used in specialized applications where paying for new silicon every few years would be too costly, which include identifying pictures of rotten food taken from an assembly line, processing cellular signals inside base stations, or prototyping custom chips. But companies like Microsoft and Baidu are increasingly using FPGAs as accelerators in data centers.
Like graphics chips, FPGAs exhibit parallelism. But unlike graphics chips, FPGAs can also be updated after manufacturing to suit machine learning algorithms, which are still evolving. That combination gives these chips benefits over custom chips from start-ups like ThinCI and Mythic, as well as Intel’s Movidius unit that target cars and security cameras.
“The inference side of machine learning is still wide open,” said Steve Mensor, vice president of marketing for Achronix Semiconductors, in a May interview. Achronix, like start-ups Flex Logix and Menta, has started to license FPGAs so that the programmable logic can be squeezed on system-on-chips (SoCs), cutting the bandwidth and power consumption that typically exempt embedded artificial intelligence.
These companies are fighting the low survival rates of programmable chip start-ups, which have had trouble keeping pace with the improvements that Xilinx and Altera get from scaling down transistors used in their chips. These firms can also invest amply in software tools for the complex task of programming FPGAs.
The fate of Tabula underlines the potential pitfalls. The company built a three-dimensional FPGA that reprogrammed itself automatically to create shortcuts between logic blocks. The architecture was supposed to enable more efficient chips that could be sold for less than $200. But the company stopped operating in 2015 after it burned around $215 million tweaking the technology and software tools.
Efinix has been trying to stay under the radar for the last five years, but it recently raised $9.5 million from investors including Xilinx. In a statement, Salil Raje, Xilinx’s senior vice president of software and IP products, said that “Efinix’s solution can address a wide variety of applications that are typically not served by today’s FPGAs.”
Cheung said that both companies would avoid stepping on each other’s toes. Efinix is trying to sell programmable chips for high-volume applications where power and size are priorities. “I have customers pre-paying before we even ship a product, which is new for this industry,” Cheung said.