The KL730 is not Kneron's first chip optimized for transformers, but it does offer significant performance improvements over previous models. It can also be clustered for larger deployments, potentially broadening its use for machine learning training. Liu is hopeful that the company's reconfigurable AI architecture, which can dynamically change the structure inside the chip to support almost any new model, will help Kneron carve out an expanded market footprint.
Key takeaways:
- Kneron, an edge AI startup, is introducing its latest KL730 Neural Processing Unit (NPU), which it claims is up to four times more energy efficient than previous models and is designed to accelerate GPT, transformer-based AI models.
- The company's silicon is primarily aimed at edge applications, including autonomous vehicles and medical and industrial applications, but it also sees potential for enterprise deployments.
- Kneron's chips use a reconfigurable AI architecture, which is different from the architecture used in a GPU. The KL730's architecture has been specifically optimized for GPT's transformer-based AI models.
- The KL730 has improved performance compared to previous generations and can be clustered for larger deployments. Kneron hopes that the ability to combine multiple KL730s will enable broader use of the technology for machine learning training.