SiMa Technologies Inc., the maker of a purpose-built machine learning system-on-a-chip, Tuesday said it has begun shipping its SiMa.ai Palette software platform for full machine learning stack development, along with two plug-in processing cards.

The San Jose, California-based company said the development platform supports the deployment of any machine learning workflow at the edge without compromising performance. Based on 16-nanometer technology, the MLSoCs include computer vision processors for image pre- and post-processing along with dedicated machine learning acceleration and high-performance application processors.

The company says its hardware achieves 10 times better performance than general-purpose chips with lower power consumption. It’s initially focusing on computer vision applications.

SiMa is one of a growing list of startups that are building hardware specifically for the development and deployment of artificial intelligence applications using machine learning and neural networks. They’ve attracted most of the $1.8 billion venture capitalists poured into chip startups last year, chasing a market that Allied Market Research expects to grow from $8 billion in 2020 to nearly $195 billion by 2030. SiMa has raised $177 million in funding, according to Crunchbase.

The SiMa.ai Palette is described as a set of scalable tools for machine learning development that deliver high performance and push-button operation. Its GStreamer workflows use advanced scripting and automation on optimized neural network models, along with pre- and post-processing functions and pipelines that permit the deployment of a single container to large numbers of edge devices simultaneously.

The new hardware includes a PCI Express half-height, half-length plug-in board and dual M.2 production boards containing purpose-built silicon for low-power edge uses. M.2 is an AI-specific high-performance parallel computation machine. Customers can use the board designs to accelerate deployment and to develop their own custom devices, the company said.

The boards include four-core Arm A65 processors, an H.264-compatible video encoder/decoder, a machine learning accelerator that delivers up to 50 trillion operations per second of performance and a computer vision processor that can be used with all popular machine learning frameworks, including TensorFlow, PyTorch, ONNX and MXNe.

The M.2 boards are priced at $599 and the PCIe boards at $749 in 10,000-unit quantities. Additional versions are being built for use in industrial temperature scenarios.

The Palette software is available in alpha test on an early access basis. The full-scale deployment will start in the second quarter.

Photo: SiMa