AI chip startup Kneron today (November 26) announced its new generation of AI chips, including the flagship KL1140. Founder and CEO Dr. Jun-Cheng Liu revealed the company's complete product roadmap for the next three years, signifying that Kneron has officially completed its full computing power layout from low-end to high-end, and announced its evolution from a simple chip supplier to a comprehensive "AI Infrastructure Company".
KL1140: Terminal fully executes the Mamba architecture, breaking cloud dependencies.
The newly released flagship KL1140 is touted as the world's first device capable of fully running the Mamba neural network architecture on terminal devices.NoteThe Mamba architecture is an edge AI chip that has recently garnered significant attention in the AI field and is considered a rising star capable of competing with the Transformer architecture.
Kneron emphasizes that the KL1140 achieves three times the energy efficiency of existing cloud solutions, while reducing costs by a factor of ten. By connecting four KL1140s in parallel, it is possible to support the operation of large language models (LLMs) with up to 120 billion parameters, while consuming only one-third of the power of traditional GPUs. This means that large language models will no longer need to rely on expensive and high-latency cloud GPUs and can be truly implemented in smart terminal devices.
Note:The Mamba neural network architecture, proposed and developed by researchers at Carnegie Mellon University and Princeton University, is a deep learning architecture based on the Structured State Space Sequence (S4) model for sequence modeling tasks. It can address the limitations of the Transformer architecture currently used for most AI operations, especially in handling long sequences. It can handle irregularly sampled data and long contexts, and maintain computational efficiency during training and inference.
Three-year blueprint revealed: covering security, general vision, and Gen AI
At the meeting, Dr. Liu Juncheng also revealed for the first time Kneron's chip plan for the next three years, covering a full range of products from high-end to low-end:
• Low-power security (KL640 / KL650):It features always-on capability and targets video security and low-power consumption scenarios.
• High-performance, cost-effective general-purpose vision systems (KL540 / KL515):Targeting common terminal devices such as IP cameras and machine vision systems.
• New generation Gen AI dedicated (KL840 / KL1150):Designed for high-end smart devices, automotive and enterprise edge servers, it boasts higher computing power and connectivity.
KNEO 350 edge server and KNEO Pi platform unveiled
In addition to the chip itself, Kneron also showcased its complete ecosystem, including the next-generation edge AI server KNEO 350 and the KNEO Pi development platform, which has accumulated 28000 developers worldwide.
Kneron points out that current AI computing models face high costs, high energy consumption, and data privacy risks. Through the KL1140 and its new series of chips, Kneron will promote its vision of "The Future Lives at the Edge," enabling applications such as smart security robots, in-vehicle systems, and enterprise private AI assistants to operate in real time without relying on cloud connections, ensuring that sensitive data remains local.
Kneron has already partnered with international companies such as Qualcomm, Hanwha, Toyota, Quanta, and Sony, and deployed its solutions in sovereign AI projects in healthcare, education, and government agencies.
