AMD announced today (August 7) that it plans to acquire Taalas, a chip startup. Taalas designs accelerators for AI inference and customizes chips for individual AI models, primarily sacrificing versatility in exchange for lower costs and higher output speeds.
Technically, the AI inference chip currently showcased by Taalas is based on Meta's Llama 3.1 model. It is built using mature (older) process technology from TSMC and uses high-speed SRAM on the chip itself.
The advantage of this approach is faster inference. The output speed of specific models is thousands of times faster than that of traditional GPUs, while costs are also lower. However, the chip is tied to a particular AI model and cannot be upgraded later.
Ljubisa Bajic, CEO of Taalas, said on the company's website that Taalas has developed “a platform that transforms any AI model into custom silicon.” He also said that after receiving a previously unsupported model, the company can implement it in hardware in just 2 months.
AMD did not disclose details about the transaction size in its blog post. However, according to public information, Taalas has raised a total of 2.19 billion US dollars in venture capital since its founding in 2023 (note: approximately 1.481 billion yuan at the current exchange rate).