Astera Labs has introduced a new high-speed switch that could change how AI systems connect and communicate. Unlike traditional options, this switch aims to be compatible with almost any kind of accelerator, offering more flexibility for building large-scale AI setups. Called Scorpio X, it packs a lot of bandwidth into a single chip, promising faster data transfer for demanding AI workloads.

What Makes Scorpio X Different

Scorpio X is a PCIe 6.0 switch that offers 320 lanes of connectivity, with a bidirectional bandwidth of 5.12 terabytes per second. This means it can handle a huge amount of data moving between different components without slowing down. In the past, PCIe switches were mainly used to connect CPUs, GPUs, and storage devices, but this new switch aims to support large AI clusters more effectively.

One of its key features is built-in compute capabilities similar to Nvidia’s NVSwitch, which helps speed up collective communications among multiple GPUs. This is especially useful for generative AI models that rely on complex, multi-part architectures. With Scorpio X, GPUs don’t have to wait as long for data, which can lead to faster AI inference and training times.

Supporting the Future of AI Hardware

Astera emphasizes that Scorpio X isn’t meant to replace Nvidia’s NVSwitch but to be a vendor-neutral alternative that works with many different systems. While Nvidia’s NVSwitch offers even higher bandwidth, Scorpio X is designed to work seamlessly with a wide range of accelerators and interconnect protocols, including PCIe and upcoming standards like NVLink Fusion.

This flexibility is important because many AI systems use a mix of hardware from different vendors. For example, some setups combine Nvidia GPUs with chips from other companies like Groq or Cerebras. Using PCIe makes it easier to connect these diverse components without needing specialized chips or redesigning hardware architectures.

Astera also plans to roll out a series of switches with varying lane counts—from 32 to 320 lanes—so data centers can choose the right level of connectivity for their needs. These switches will work with a management platform called COSMOS, which helps monitor and troubleshoot network issues, making deployment smoother. The new switches are expected to start sampling later this year, with broad availability by the second half of 2026.

Overall, Astera’s new PCIe switch aims to give AI developers more options and better performance. By supporting a broad ecosystem of hardware, it could make building and scaling AI systems simpler and more efficient.

Inspired by

• https://go.theregister.com/feed/www.theregister.com/2026/05/05/nvlink_astera_big_swich/

Sources

• asteralabs.com