For AI to be able to operate in a physical environment, it requires extremely efficient embedded software and latency-free radio traffic in the background, whether it is wifi, 5G or the 6G in the future.
Unikie has been at the heart of physical AI since its founding in 2015.
“Since the beginning, we have focused on embedded systems, and that’s really all we’ve done. At the same time, all the work we’ve done has paved the way for the emergence of physical AI,” says Seppo Kolari, one of Unikie’s founders and the head of Unikie US.
According to Kolari, in-depth expertise in embedded systems is the reason why Unikie, which originated in Tampere, has grown into an international success.
Physical AI requires experts from the Wild West
So why is the development of embedded systems so demanding that it separates the wheat from the chaff? According to Kolari, the reason lies in the merciless limitations of the physical world.
“The embedded side is a kind of Wild West, where there are no ready-made standards. There are a huge number of different device environments, each with its own strict limitations: limited memory, strict power consumption requirements, and performance limits,” Kolari describes.
These physical constraints – space, memory, even weather conditions – determine how software can be built. In addition, physical AI emphasizes absolute real-time requirements and security.
“If artificial intelligence hallucinates while making a PowerPoint presentation, it’s not dangerous. But messing around with the systems of a car or other heavy machinery can, at worst, affect human lives,” says Kolari.
When the great is made even better
Kolari is responsible for finding customers for Unikie in the world’s toughest market, the United States. Physical AI solutions are often built on top of services and products that have already been built by world-class experts. With the help of physical AI, a brilliant can be made even more efficient or accurate.
“In the spring, we launched UnikieMind, an agent-based artificial intelligence platform to accelerate software development, which allows us to offer our customers significant productivity leaps in embedded development. Using the same platform, we can build a prototype for a customer meeting, for example, which multiplies the speed of image processing. Instead of waiting for the results for three minutes, the doctor receives them in ten seconds,” Kolari concretizes.
GPU optimization for superior imaging accuracy and speed
Unikie’s exceptional ability to optimize the interaction between hardware and software is reflected in customer projects. Unikie’s Principal AI Engineer Heikki Lehtosalo uses Grundium as an example, whose microscope scanners are trusted tools by pathology, laboratory science and research professionals around the world.
“We did GPU optimization for them, i.e. to improve the efficiency of the graphics processor in a situation where the microscope slides over the preparation and takes thousands of images. The software seamlessly combines these images and removes noise so that the end result is a perfectly sharp and uniform image,” Lehtosalo says.
“The customer was positively surprised by the level of performance we reached.”
5G signal decoder latency reduced by 80 percent
Further evidence of Unikie’s optimization expertise comes from 5G networks, including physical-layer signal processing and coding on NVIDIA’s AI Aerial platform.
“We managed to reduce the latency of processing by 25 per cent and up to 80 per cent for different algorithms. This is an excellent result,” Lehtosalo says with satisfaction.
Seppo Kolari highlights the true scale of the achievement.
“You have to take into account that in this case, we are talking about code that has already been optimized by world-class developers, but its operation could still be significantly improved.”
According to Lehtosalo, in order to optimize the software to the extreme, the coder must combine an understanding of the algorithm with an in-depth knowledge of how the hardware works, taking into account the context of the entire system.
“GPU optimization in particular has a lot of moving parts: how to increase parallelism and at the expense of what, how to use the memory hierarchy efficiently, how to reduce synchronization between threads, and much more,” Lehtosalo lists.
Lehtosalo says that in GPU computing the common approach is to hide the latency of operations through parallelism, utilizing an efficient hardware-level scheduler, to achieve a higher throughput.
“This is what the GPU is designed for, but there are particularly interesting challenges when this approach doesn’t work, as is often the case with 5G, for example. Instead of maximizing throughput, you need to keep the latency of the entire pipe below a strict limit in all cases.”
The content of Lehtosalo’s work changed dramatically when he started to develop an AI agent system that implements and optimizes algorithms instead of doing that manually. Today, the agent system does most of the low-level optimization, but according to Lehtosalo, a human expert is still needed to determine what exactly is to be optimized.
“You get what you measure, and it’s not always clear what you should measure to achieve the desired result,” he sums up.
Whether it is speeding up cancer diagnosis or developing a latency-free 5G network, fine-tuning the code is not a goal or achievement in itself. Essentially, it is about creating significant business benefits and competitive advantage for customers.
Read more about Optimind: Software Performance Optimization
This article was originally published in Finnish in Tekniikka & Talous.
