SiMa.ai’s Machine Learning System-on-Chip (MLSoC) is a software-centric ML chip that targets the edge
SiMa.ai, a top-class machine learning company, has just launched a software-centric purpose-built machine learning chip platform which provides a high-performance, effortless machine learning deployment for embedded edge applications: the MLSoC.
The MLSoC features, amongst others, a dedicated machine learning accelerator, computer vision processors, high-performance application processors, memory and communication interfaces, and a secure system management unit.
The chip can be used as a stand-alone edge-based system controller or as a machine learning offload accelerator for processors, ASICs, and other devices. It has a very low power consumption rate but delivers high machine learning processing capacity.
Asides from the fact that the MLSoC supports legacy applications as well as future machine learning use cases with high performance, low power, and safe and secure machine learning inferencing, the chip also has the ability to run up to four distinct DNN models concurrently. It also supports and optimizes DNN models from a variety of frameworks ranging from PyTorch, TensorFlow, MXNet, ONNX, etc.
Features of the MLSoC include:
- Machine Learning Accelerator – Capable of pushing 50 TOPS for neural network computation at 10 TOPS/W
- Computer Vision Unit – Has a 4-core Synopsys ARC EV74 video processor with up to 600 16-bit GOPS
- DRAM interface – 4x 32-bit LPDDR4 memory controllers and PHY interfaces at up to 3733 MegaTransfers per second.
- High-Speed I/O Subsystem – 4x 1GbEthernet ports and 1x PCIe Gen4 8-lane interface
- Low-speed I/O Subsystem – 2x SPI buses, 2x I2C buses, 32 GPIO pins and SDIO/eMMC flash.
- Application Processing Unit – a group of 4x Arm Cortex-A65 dual threaded processors running @ up to 1.15 GHz
- Video Encoder/Decoder – Supports the H.264/H.265 decode at 4k60 and H.264 encode at 4k30
- Boot and Secure Unit – Provides secure key storage in an OTP memory as well as key management.
SiMa.ai claims that the machine learning chip is about 10 times better than its alternatives even at the lowest power, with the ability to run any computer vision application, any network, any model and any framework, while giving push-button results.
“When we started SiMa.ai 3.5 years ago, we set out to deliver a disruptive 10x performance improvement over alternatives and provide a scalable industry-leading ML experience solving computer vision applications,” says CEO Krishna Rangasayee. “Today, we are delighting customers by delivering on that promise and exceeding their expectations. We are excited to take our very first purpose-built software-centric MLSoC to volume production.”
The MLSoC is applicable in smart vision, autonomous vehicles, drones, healthcare and Robotics and Industry 4.0.
In addition to this, the company launched an evaluation board that places the chip on a PCI Express card with 16 GB RAM and eMMC storage. This is operable as an accelerator card in a host machine or stand-alone SBC with 4x GbE ports.
The company has not yet made any public details regarding the price for the MLSoC and evaluation board but other useful details can be found on their website. There’s also a form on the website that you will be required to fill in case you want direct access to them.
