Untether AI joins UCIe Consortium to propel chiplet technology
Untether AI has recently joined the UCIe Consortium, an industry organization that has developed an open specification for connecting chiplets within a package.
In recent years, the landscape of compute workloads has undergone significant transformations, primarily driven by the surge in AI workloads. As the complexity of AI and software algorithms continues to evolve, there has been a corresponding increase in the demand for heterogeneous computing utilizing domain-specific architectures.
Arun Iyengar, the CEO of Untether AI, mentions that using chiplet technology and UCIe’s universal standards, Untether AI plans to provide energy-efficient AI acceleration chiplets to the market.
“We are already seeing great acceptance of UCIe from our customers and are excited to see the proliferation of products with UCIe,” adds Iyengar.
To achieve optimal efficiency, companies are now leveraging a combination of AI accelerators, CPUs, GPUs, memory and networking technologies. However, this approach presents unique challenges in energy consumption and data transport, especially as neural network models continue to grow.
UCIe offers a universal standard facilitating efficient communication between components, addressing challenges in die-to-die interconnects. The latest release, UCIe 1.1, further expands its support to encompass automotive applications, specifically autonomous vehicles, which remains a prominent focus for Untether AI.
Untether AI notes that its chiplets enable the creation of compact, high-performance compute systems. The technology further allows the development of tailored chipsets to meet specific application requirements.
“Untether AI’s contribution will help showcase the benefits of UCIe, a low-power, high-speed, and open industry standard chiplet interface,” states Dr. Debendra Das Sharma, the UCIe Consortium chairperson.
Untether AI says it offers energy-centric AI inference acceleration from the edge to the cloud and supports all neural network models. According to the company, its at-memory compute architecture eliminates data movement bottlenecks and offers high-performance, low-latency neural network inference without compromising accuracy.
Founded in 2018, Untether AI is backed by CPPIB, GM Ventures, Intel Capital, Radical Ventures and Tracker Capital.
Eta Compute, NXP Semiconductors team up to optimize edge AI development
Comments