Assuming it is the software that enables or prevents the adoption of a new AI chip, a reasonable assumption, SiMa has created a new tool called Palette Edgematic to simplify AI solution creation greatly. And build demand for their silicon.
As everyone says, “It’s the software, stupid!” when it comes to porting AI models to a new piece of hardware. A vast repository of tools from NVIDIA would need rewriting and replacing in most software stacks, especially if you want to get the maximum performance out of your chip. That is a lot of challenging work.
But with Large Language Models (LLMs), the community has done much of the work for you, at least to a first level of performance. Hugging Face, for example, has hundreds of open-source LLMs you can use free or at a smallish charge. And for makers of new hardware, like Cerebras or SiMa.ai, this provides a far better starting point.
Building AI Workflows
But most AI applications aren’t just one model. They are interconnected AI models, each performing a different task. SiMa.ai, an edge AI provider, has developed a dreamboat solution that brings a drag-and-drop interface to creating AI workflows.
Say you want to build an edge AI app to run Centernet, an object detection AI that detects each object as a triplet, rather than a pair, of keypoints, improving precision and recall with the new SiMa.ai Palette platform. As shown below, you can create a workflow from pre-processing, processing, and post-processing with a simple point-and-click interface.
A screen shot from the Palette Edgematic platform. SIMA.AI
Now you can take this new object, Centernet, and crop, resize, and encode the output with more cdrag and drops. Simple! Even I could do it! Now you can connect any hardware, such as —surprise— the SiMa MLSoC and click to say “Run on this device,” and you are done.
Conclusions
It was really refreshing to get briefed on this by SiMa.ai. They were justifiably excited by the opportunity to make it easier for customers to adopt their solution (chip + SW) without a learning curve or even a sales call. Execution of high-performance software is not always the case with semi and edge companies, but with the introduction of Palette Edgematic SiMa is delivering on its promise to put software first. Plus, it comes on the heels of the MLCommons® ML Perf 3.1 benchmark results, where, notably, SiMa.ai scored higher on performance and power efficiency than its primary competitor (Nvidia OrinX and AGX) in the Closed Edge power category for the second time this year.
As their CEO, Krishna Ragnasayee, said, “Push Button… Wow!” And just like that, the company expects its customers to be able to cut embedded-edge ML development processes from months to minutes.
Now, that’s how to generate demand for a new hardware platform! If you want to try the new software, it is available for free here.