SiMa.ai adds boards, software to its edge AI/ML pitch

As artificial intelligence and machine learning are moving to the edge, and in some cases into endpoint devices, a throng of start-ups have emerged with solutions designed to accelerate what existing system-on-a-chip platforms can do for specific AL and ML applications. Not many of them are trying to be the next Nvidia, AMD, or Intel, the companies behind those SoCs, but that appears to be what SiMa.ai has in mind.

The San Jose, California, company, which as of late last year had boosted its total venture funding to $187 million, was founded by Krishna Rangasayee, a veteran of Xilinx, which itself is now part of AMD.

Rangasayee believes that the CPUs, GPUs, DSPs that such legacy semiconductor giants are providing are not meeting the needs of the modern ML movement–and his company has developed its own ML SoC to prove the point. Most recently, SiMa.ai this week announced availability of two new PCIe-based production boards that it said will scale embedded edge ML deployments for “key customers.” It also announced its Palette software platform to provide “a pushbutton experience” to developing complete end-to-end ML applications leveraging the heterogeneous SiMa.ai ML SoC.

“The vast majority of the embedded edge customers want ML but have not ventured into ML-based solutions, primarily because they miss internal capabilities and/or a platform that supports legacy computer vision and new machine learning frameworks,” Rangasayee told Fierce Electronics via email. “The customers that have tried ML are challenged by large companies leveraging legacy technologies like GPUs, CPUs, and DSPs that do not meet the performance and power needs of the embedded edge customers—and not having a unified software suite that takes through the product life cycle.”

SiMa.ai’s “software-centric” approach to its ML SoC changes the game, delivering 10X power efficiency of alternatives, the company claimed. SiMa.ai did not identify the “key customers” for its new production boards, but has said in the past that it is working with customers in areas such as robotics, autonomous driving, security cameras, and more.

SiMa.ai is not the only edge AI/ML start-up in town, however, and new companies like Perceive, Quadric, Flex Logic, Hailo Technologies, and others also have jumped into the market to meet what they see as a growing need. But some of those companies are not focused on SoCs, but AI acceleration techniques, an important distinction, according to Rangasayee. 

“The new ML startups just accelerate ML, which is just a portion of the problem and not solving the entire computer vision end-to-end pipeline,” he said. “SiMa.ai’s MLSoC is the first purpose-built software-centric solution that addresses the challenges customers have faced in scaling ML and is focused on effortlessly scaling ML at the embedded edge.”

Karl Freund, founder and principal analyst of Cambrian-AI Research, also noted the difference in approach in a recent bit of research on Perceive’s Ergo2 product update.

He wrote, “There are many companies readying or shipping chips for Edge AI, including SiMa.ai, Hailo Technologies, AlphaICs, Recogni, EdgeCortix, Flex Logix, Roviero, BrainChip, Syntiant, Untether AI, Expedera, Deep AI, Andes, Plumerai, in addition to Intel, AMD (Xilinx) and of course NVIDIA. Some, like NVIDIA and SiMa.ai are heading down the SoC route, where the chip offers a more complete solution including Arm or RISC-V CPU cores and I/O. In contrast, Perceive (and others such as Hailo) has focussed on customers who are looking for an AI accelerator that attaches to an SoC for a specific application.”

Jack Gold, president and principal analyst at J. Gold Associates, said the challengers face a difficult fight in the market. “AI/ML techniques are not really standardized, with a wide variety of algorithms. So many of these companies optimize their [hardware] for key algorithms and then claim they are the best,” he said. “But there is still a pretty general need for GPU programmability, which is why Nvidia is still a central component, as it employs a variety of techniques and supports many algorithms, with many more programmers using it.”

He added that “the issue for the startups is, can they build enough momentum to make a dent in the market? And will they reach critical mass with customers? My guess is most of them won’t survive long term, although some may be acquired which is the play many of them, and their VCs are hoping from (like Intel acquired Habana Labs a few years back). Will they outcompete the big semi guys, and the folks like Google, AWS and Microsoft for their cloud instances of AI/ML? It’s a long uphill battle.”