SiMa.ai debuts new system-on-chip for multimodal AI at the network edge
Artificial intelligence chip startup SiMa Technologies Inc. has unveiled what it says is the first system-on-chip platform that's specifically designed to handle multimodal AI workloads at the embedded edge.
With today's launch, the startup reckons it can power a wide range of AI workloads in industrial robots and other edge computing devices with much lower latency than was previously possible.
The new chip, called the MLSoC Modalix, supports various kinds of AI models, including large language models such as OpenAI's GPT family, convolutional neural networks, transformer models and more.
According to the startup, AI has already had a huge impact on the way humans work together with machines, powering edge applications in warehouses, factory floors, weather stations, autonomous cars, drones and other areas. However, most of these edge applications run much smaller models than the ones it's targeting with its latest chip.
The company is specifically aiming at so-called multimodal AI applications, which can understand and process multiple inputs in the shape of text, images, audio and more. Such models typically require much more processing power. Until now, they've always relied on cloud-based processors, resulting in lower latency when used in edge applications.
That will change with the MLSoC Modalix chip, as it packs much more processing power than the company's original MLSoC platform. According to the startup, the MLSoC Modalix will be made available in a number of configurations, including 25, 50, 100 and 200 TOPS - a metric for performance, where one TOP corresponds to 1 trillion computations per second. The chips are integrated with a 25-megabyte memory pool, which allows them to store data directly on the chip without using external random access memory.
SiMa.ai said the new chip is built on Taiwan Semiconductor Manufacturing Co.'s six-nanometer process and packages four vector processing units for running computer vision algorithms, as well as modules for video encoding and computer vision tasks. There's also an integrated ISP module that allows the chip to interface with Mobile Industry Processor Interface-based cameras and Ethernet-connected cameras, as well as four lanes of PCIe Gen 5 for data ingress and egress to and from the chip.
Meanwhile, general-purpose computing tasks are offloaded to an Arm-designed eight-core Cortex A65 central processing unit that runs Yocto Linux, an operating system specifically designed to run connected edge devices.
SiMa.ai says MLSoC Modalix will enable developers to "push the envelope" in terms of edge AI performance, running end-to-end generative AI application pipelines on just a single chip. The result will be far more powerful edge AI applications, including drones that can operate autonomously to perform tasks in agriculture, industrial maintenance and building inspections, and autonomous vehicles that can drive themselves while performing routine maintenance checks on themselves. Other use cases include things such as medical imaging and robot-assisted surgery in healthcare, and smart retail applications, the company said.
Besides the new chip, developers also get access to SiMa.ai's suite of low-code AI development tools, known as Pallet, which includes a compiler that enables AI models to run more efficiently on the hardware, plus a layered, direct proactive data prefetch capability that ensures the underlying models have access to data in real time.
SiMa.ai founder and Chief Executive Krishna Rangasayee said his company has already seen significant uptake of its first-generation MLSoC chip. "With the introduction of MLSoC Modalix, SiMa.ai now provides coverage from convolutional neural networks to generative AI and everything in between, with industry-leading performance and power efficiency," he said.