[AI Revolution](101) Ultra-large AI, Light and Fast... SqueezeBits That Even Intel Loves

Lightweight Image AI Drawing in 7 Seconds
Samsung and Naver Invest... Lightweight Toolkit Released
"Experience in Both HW and SW... Will Bridge the AI Ecosystem"

Kim Hyung-jun, CEO of AI-specialized startup SqueezeBits, visited Phoenix, Arizona, USA last week for 'Intel Vision 2024.' This was to jointly research software (SW) based on Intel's 'Gaudi' AI chip, following the AI chip alliance between Intel and Naver. Lightweighting is considered a core technology that connects the AI value chain by finding the optimal match between AI models and the semiconductors that run them. This is why, despite being a startup in its third year, SqueezeBits has received investments from leading domestic AI players such as Samsung, Naver, and Kakao.

Kim Hyung-jun, CEO of Squeeze Beats. Photo by Heo Young-han younghan@

CEO Kim is a founder with a research background. During his master's and doctoral studies at POSTECH, he researched AI semiconductor design and lightweighting technology. Due to the nature of his research field, he often collaborated with companies through industry-academia projects. Each time, he personally felt how much the industry was thirsty for lightweighting technology. To address this demand in the industrial field, he and his research colleagues founded SqueezeBits in 2022.

SqueezeBits' core technology is AI compression through quantization. It simplifies 32-bit data into lighter 8-bit or 4-bit units while maintaining model performance. Simply put, in calculations multiplying 8-digit numbers, rounding the numbers and keeping only the leading digit speeds up the computation significantly. CEO Kim explained, "Our strength lies in making the model lighter by discarding less important information without degrading AI model performance, while maintaining performance as much as possible."

Using this method, they succeeded in running the AI model 'Stable Diffusion' quickly on Android smartphones. Stable Diffusion is an AI that converts text into images and is a super-large model with over one billion parameters. The larger the number of parameters, the better the AI performance, but it requires substantial infrastructure to run. Therefore, it is mainly operated on graphic processing units (GPUs) in cloud environments. SqueezeBits lightweighted it to run on smartphones without cloud connection. While maintaining performance such as image quality, the model was lightened to generate images in 7 seconds. This is faster than Qualcomm’s Snapdragon 8 Gen 2 chip, which took 15 seconds to run Stable Diffusion in February last year.

Kim Hyung-jun, CEO of Squeeze Beats. Photo by Heo Young-han younghan@

To commercialize the technology, they also launched 'Owlite' last March. It is an AI lightweighting toolkit that applies SqueezeBits' know-how in the form of Software as a Service (SaaS). According to the company, clients can improve AI inference speed by an average of 3.6 times and reduce memory usage by an average of 4 times using this toolkit. CEO Kim said, "Many AI startups in Silicon Valley are running at a loss due to GPU costs," adding, "Startups respond well because they can quickly update AI models and accumulate lightweighting know-how."

Since most companies use NVIDIA GPUs, OwLite is optimized accordingly. In the future, they plan to expand to Intel central processing units (CPUs) and domestic semiconductor companies’ neural processing units (NPUs) as well. This is to provide clients with various options. CEO Kim emphasized, "Since we have experience in both semiconductor and software design, we want to serve as a bridge between the two," adding, "Through this, we will help lower infrastructure costs and support the creation of sustainable AI."