Ergo2 is up to four times faster than Perceive’s first-generation Ergo chip, and the company claims it can handle large models like NLP.
Edge AI is coming into its own, releasing a variety of chips that offer low cost, low power, and high performance. As training AI models gain more attention in the media, inferential processing will end up gaining the majority of revenue, especially at the edge. Global Market Insights, a respected market intelligence firm, expects the market for Edge AI to reach over $5B by 2023, with a CAGR of 20% over the next decade. While our gut feeling is that $5B is too high, we feel the 20% growth forecast is too low.
In any case, the market attracts many competitors, including Perceive, developed in 2018 from Xperi Corporation, to focus on this opportunity, and now its second product is already ready for the market.
What did Persive declare?
The company isn’t replacing its Ergo product, but rather adding a higher-performance, more capable chip for demanding edge applications. As the table below shows, the new device provides a huge boost in image classification and consumes less than 20 mW. That’s milliwatts, or 1/1000 of a watt. We know of no competitor that can claim that and still deliver nearly 1000 guesses per second.
The new Ergo 2 offers four to five times the image classification performance of its predecessor … [+]
There are several companies making 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, and Plumerai. Intel, AMD (Xilinx), and of course NVIDIA. Some, like NVIDIA and SiMa.ai, go the SoC route, where the chip offers a more complete solution, including Arm or RISC-V CPU cores and I/O.
Both Perceive products demonstrate excellent power efficiency, although we note its competitors … [+]
In contrast, Perceive (like Hilo) is focused on customers looking for an AI accelerator that fits into an SoC for a specific application. Interestingly, the Ergo chip does not require external DRAM, although it does support connectivity to NOR flash to accommodate the load for larger models. This can be a cost advantage for applications such as speech-to-text and audio applications, video processing tasks such as video super-resolution and pose detection. Compared to existing products such as the Hailo-8 accelerator at 2-4 watts, Ergo aims for lower power (tens of milliwatts vs 2-4 watts for Hailo-8) albeit at a reduced rate.
Perceive is designed to connect to a device application processor SoC.
Conclusions
As we’ve always said, penetrating the edge market is much easier than practicing in the data center because different applications have dramatically different requirements. While an image processor for autonomous vehicles requires high performance at high power levels, an embedded speech to a smart camera or text processor requires a low power envelope and low cost. Many auto manufacturers will choose an SoC like NVIDIA Drive rather than designing their own SoC; Tesla is an exception to the rule.
As a result, there’s plenty of room for specialized low-power processors like Perceive, and the company is smart to quickly expand their first push with a faster sub-watt processor. Software that enables larger models to work efficiently on the new Ergo2 will be key to their success.
We’ll note that there are many more competitors entering this market, so check back soon to stay up-to-date!
Disclosures: This article expresses the views of the author and should not be construed as advice to buy or invest in the companies mentioned. My firm, Cambrian-AI Research, is fortunate to have many semiconductor firms as our clients, including NVIDIA, Intel, IBM, Qualcomm, Esperanto, Graphcore, SImA, ai, Synopsys, Cerebras Systems, Tenstorrent, and Ventana Microsystems. We have no investment positions in any of the companies mentioned in this article. For more information, please visit our website https://cambrian-AI.com.
