Alberto Romero, Cambrian AI analyst, contributed to this article
SiMa.ai is a machine learning startup founded in November 2018 with a mission to transform and ignite the embedded Machine Learning (ML) edge market by addressing the shortcomings of current solutions. Now, with $150 million in funding behind it, SiMa.ai has claimed to have created the first software-centric “MLSoC” (Machine Learning System on Chip) platform capable of running edge vision models at 10 times better. energy efficiency than its competitors. Rather than going it alone, the company partnered with leading silicon IP vendors, including Synopsys and Arm, to complement their proprietary ML IP to accelerate the entire workflow while speeding up time to market. .
A software-centric, purpose-built solution for ML on the edge
The company has announced that it is launching and shipping the MLSoC platform to customers. This is the first step towards achieving its ambitious goal of reshaping the embedded edge market – which the company estimates at an annual SAM of an estimated $1 trillion. By combining its proprietary technology with state-of-the-art hardware from partners, SiMa.ai claims it can outperform competitors who tend to trade easy-to-use software and energy efficiency for high-end architecture.
While others have also created specific software and hardware designs, SiMa.ai has taken this approach to the extreme. Developers want easy-to-use, energy-efficient hardware that they can use while continuing to leverage and maintain legacy applications. That’s what SiMa.ai’s MLSoC platform is all about: a software-centric (flexible and easy to use), purpose-built (not adapted to, but designed for, the embedded edge) platform capable of meeting the demands of edge intelligence.
This approach allows the company to target a wide range of markets, including smart vision, robotics, healthcare, drones, government and autonomous vehicles. The MLSoC platform allows SiMa.ai to meet the requirements of embedded edge devices that all share a characteristic feature: they have size, weight and power limitations (SWaP profiles). Edge solutions adapted from other markets have traditionally struggled to adapt to these strict constraints.
computer vision with 10 times the energy efficiency
SiMa.ai claims that its MLSoC solution can handle any framework, model, and computer vision application with the Apache TVM-based Front-End compiler and the optimized SiMa.ai Back-End. The company states that the MLSoC platform offers 10x better energy efficiency (FPS/W) over alternatives in each framework (PyTorch, ONNX, TensorFlow, TFLite, etc.) on more than 120 neural networks tested. The co-design of hardware and software enables compute scheduling and data movement that is not possible with data center hardware designs that are easily adapted for embedded edge tasks.
SiMa.ai claims that its software can make any computer vision model ready for execution on the MLSoC platform in minutes instead of months. SiMa.ai’s software stack automatically plans and optimally distributes code across MLSoC subsystems, such as the Synopsys-provided EV74 Vision Processor, without manual intervention — a one-button experience.
The Secret Sauce of SiMa.ai: A Holistic Vision Implemented in a SoC for ML
There is a reason why SiMa.ai has managed to deliver a highly competitive solution to disrupt the embedded edge market. The company’s vision is reflected in the holistic optimization of MLSoC. It is designed not only to speed up the ML inference elements, but to run the entire pipeline, including the pre- and post-processing phases of the application, which are often overlooked. This has enabled SiMa.ai to amply meet and exceed customer demands.
The holistic approach of SiMa.ai is also reflected in the architecture of the MLSoC Platform. It includes an ML accelerator (MLA), a tile-based element that achieves 50 TOPS at 5W, and takes care of the ML calculations. It is accompanied by a computer vision unit made up of four-core Synopsys ARC EV74 processors. A number of other important functions are handled by an application processing unit (4x Arm Cortex-A65) and a combined video encoder-decoder. Different stages of the ML application are solved by different elements within SiMa.ai’s MLSoC platform.
The software development kit (SDK) provides tools to automatically quantify the models and optimize performance, while minimizing latency at the touch of a button, saving customers time and money.
SiMa.ai’s is essentially a software company that builds its own silicon. It has created a software-centric, purpose-built, flexible, energy-efficient, scalable, easy-to-use MLSoC platform that could overcome the limitations of increasingly outdated silicon-first solutions. The company delivers 1/10 the response time at 1/10 the power compared to competitors. SiMa.ai is built on a strong foundation and the first generation product seems to be clearly differentiated in a crowded market.
disclosures: This article reflects the views of the authors and should not be construed as advice to buy from or invest in the companies mentioned. Cambrian AI Research is fortunate to have many, if not most, semiconductor companies as our customers, including Blaize, Cerebras, D-Matrix, Esperanto, FuriosaAI, Graphcore, GML, IBM, Intel, Mythic, NVIDIA, Qualcomm Technologies, Si- Five, SiMa.ai, Synopsys and Tenstorrent. We have no investment positions in any of the companies mentioned in this article and do not plan to start one in the near future. For more information, visit our website at https://cambrian-AI.com.