US startup Baya Systems has made a significant breakthrough in AI chiplet technology with the development of a chip interconnect that promises to boost performance by a factor of 100. The NeuraScale switch fabric technology, created by Baya Systems, aims to address critical scaling and data movement challenges in AI infrastructure by offering a 100x increase in node density for the UALink standard.
This advancement is crucial for AI workloads that require enhanced performance, throughput, and efficiency, especially as traditional crossbar switching architectures struggle to keep up with the demands of modern AI applications.
The NeuraScale switch fabric represents a revolutionary approach, supporting 256 ports per chiplet with a throughput of 1 Tb/s. This enables next-generation AI infrastructure to achieve high-speed switching at over 2GHz in a 4 nm process technology, with minimal latency and efficient data transfer capabilities.
Moreover, the architecture of NeuraScale is modular, simplifying the physical implementation for each SoC while facilitating seamless scaling across multiple chiplets. It is compatible with current ARM AMBA interconnect protocols and is designed to meet future standards such as UALink, UCIe, and Ultra Ethernet compliance.
Dr. Sailesh Kumar, CEO of Baya Systems, emphasized the necessity for advanced interconnect solutions in meeting the evolving demands of AI workloads. He stated, “NeuraScale empowers a radical growth in node density, providing energy-efficient and area-efficient switching solutions for next-gen AI. With our platform, we enable unprecedented performance and scale, while significantly reducing implementation challenges and time to market."
- Baya Systems raises $36m for chiplet technology
- Startup emerges from stealth with chiplet interconnect
- Arteris supports AI chip design with tiling mesh network IP
- AI tool boosts chip interconnect
He also highlighted the integration of NeuraScale with the WeaverPro software-driven interconnect development platform, which accelerates design, analysis, and optimization processes. This integration streamlines development cycles, simplifies large-scale interconnect implementation, and overcomes bottlenecks associated with traditional crossbar architectures.
Karl Freund, founder of Cambrian-AI Research, acknowledged the significance of innovative interconnect technologies in scaling AI applications. He noted, “Nvidia has demonstrated the benefits of high-performance networking for scaling AI with its NVLink proprietary interconnect. Now, the industry is moving towards the UALink standard, which could democratize scaling, creating opportunities for companies like Baya Systems and their fabric solutions."
Furthermore, Peter Onufryk, President of the UALink Consortium, highlighted the collaborative efforts within the industry to drive innovation and address the challenges of AI scaling. He mentioned, “The Ultra Accelerator Link Consortium (UALink) aims to create an open ecosystem that fosters innovation and collaboration in scale-up. Our first specification will support up to 1K accelerators in an AI pod, welcoming companies like Baya Systems to meet the demands of AI scaling."
Baya Systems has established partnerships with key industry players to implement NeuraScale in chip designs, with plans to make the technology available to a wider ecosystem by Q2 2025. Notably, Tenstorrent is already leveraging Baya technology for an AI chiplet design based on RISC-V architecture.