Perceive Challenges Qualcomm and Ambarella and Crafts A Niche with Power Efficient Multi-Modal Edge AI Processor

In January 2023, Perceive announced the launch of Ergo 2, the company’s new Machine Learning (ML) processor. Compared to the company’s first ML processor, Ergo, launched in 2020, Perceive claims Ergo 2 offers 3X to 4X the performance of Ergo in a sub-100 Milliwatt (mW) power envelope and same-size package.

Ergo 2 can support advanced Deep Learning (DL) neural networks, such as YOLOv5, RoBERTa, Generative Adversarial Networks (GANs), U-Nets, and transformers. This allows for higher-resolution sensors and a more comprehensive range of applications, including video processing tasks, such as video super-resolution and pose detection; audio applications, such as acoustic echo cancellation and event detection; and language processing applications, such as speech-to-text and sentence completion.

At the same time, Perceive also provides a comprehensive suite of tools for ML application developers, including a model zoo with example networks and applications, a neural network compression and optimization toolchain, and a Software Development Kit (SDK) for building and deploying embedded applications using Ergo 2.

The upgrade from Ergo to Ergo 2 allows Ergo to move up the performance curve, add support for larger models, such as transformer networks, and run multiple heterogeneous networks simultaneously. The capability to offer higher Artificial Intelligence (AI) performance with ultra-high power efficiency allows Ergo 2 to support multi-modal applications, making it attractive to more enterprise-focused devices, in addition to the consumer device and smart home camera markets targeted previously. The company is moving to enterprise and retail cameras, high-end wearables, and laptops with more complicated use cases.

Now targeting the wider on-device AI compute market, Perceive competes directly against more established players like Qualcomm, MediaTek, Ambarella, and Amlogic, as well as other innovative AI startups, such as Gyrfalcon Technology in the on-device computer vision space and GrAI Matter Labs in on-device audio. Nonetheless, Perceive has successfully carved out a market niche for itself, similar to what Hailo and Syntiant have achieved in their respective market segments, namely gateways for Hailo and TinyML for Syntiant.

By pushing the envelope on power efficiency, Perceive is an ideal partner for device manufacturers and application developers constrained by battery power or Universal Serial Bus (USB) power, or those that have currently maximized the computing resources of their devices’ Central Processing Unit (CPU), but are not looking to make significant trade-offs in terms of performance, power efficiency, form factor, and Bill of Materials (BOM) cost. Perceive also announced that Ergo 2 will coexist in the market with Ergo. This allows the company to offer a tier-based product portfolio that addresses a wide range of applications, similar to Syntiant’s approach with its Neural Decision Processor (NDP) product line.

To push Perceive to greater heights, ABI Research would like to offer two recommendations. First, one way for Perceive to gain more market influence is by providing dedicated services for specific or multi-modal applications. These services will be able to provide tight integration and optimization of hardware and software, and can take full advantage of the chipset capabilities. Developers can then focus on model design and development without worrying about other operational parameters, such as AI model performance and power consumption.

The other suggestion is to consider supporting open-standard programming models, such as SYCL. Designed to be agnostic to all chipsets, the SYCL programming model allows developers to write highly parametrized kernels that can then be tuned for devices powered by heterogeneous architectures, including CPUs, Graphic Processing Units (GPUs), Field Programmable Gated Arrays (FPGAs), and custom AI accelerators. Enabling this would allow Perceive to tap into a broad end-user community that aims to avoid vendor lock-in with development language and framework.