AMD Unveils Helios MI455X Platform with UALink-over-Ethernet Interconnect

AMD's new Helios MI455X AI platform, showcased at Computex 2026, will initially deploy UALink-over-Ethernet. While this leverages existing infrastructure, it introduces latency and overhead trade-offs for scale-up AI training workloads.

Chips & Climate Desk
Updated Nov 5, 2026 2 min read
AMD Unveils Helios MI455X Platform with UALink-over-Ethernet Interconnect

AMD has unveiled its next-generation Helios rack-scale AI platform, featuring EPYC ‘Venice’ processors and Instinct MI455X AI accelerators. Partners displayed these units at Computex 2026 in Taipei, Taiwan, with availability expected later this year. The platform aims to rival Nvidia's NVL72 VR200 machines. A key architectural choice for these initial Helios systems is the use of UALink-over-Ethernet for scale-up connectivity, a decision that may impact performance in specific demanding workloads.

Each Helios unit is configured with 72 AMD Instinct MI455X accelerators, providing a total of 31 TB of HBM4 memory and an aggregate bandwidth of 1400 TB/s. AMD estimates the compute performance at approximately 2900 FP4 dense PFLOPS. This configuration positions Helios favorably against Nvidia's VR200 in memory capacity, potentially offering an advantage for memory-intensive tasks like running large language models, according to Tom's Hardware.

The Helios platform supports both UALink and UALink-over-Ethernet interconnects, but early systems will utilize the latter. This approach allows AMD to leverage the existing, widely supported ecosystem of validated Ethernet components, accelerating deployment for hyperscalers and cloud providers. However, 'true' UALink interconnects, which are still pending validation by AMD's AI customers, are slated for later availability.

The decision to run UALink, an open standard for direct memory access and cache-coherent communication, over standard Ethernet transport introduces inherent limitations. Unlike dedicated scale-up fabrics, Ethernet was designed for general-purpose networking. This encapsulation can lead to higher latency, increased protocol overhead, and less deterministic performance, which is critical for large AI training jobs requiring synchronized operation across all accelerators.

For large-scale AI training, where hundreds of gigabytes of model parameters must be synchronized across accelerators in microseconds, communication efficiency is paramount. If the UALink-over-Ethernet interconnect cannot efficiently supply data to the accelerators, some of the theoretical compute performance of the MI455X hardware may not be realized in practice. This contrasts with Nvidia's NVL72 VR200, which employs a fully integrated, proprietary interconnect.

Helios systems will also incorporate Pensando Vulcano network interface cards, among the first 800 GbE NICs compliant with the Ultra Ethernet specification. These NICs offer up to 43 TB/s of scale-out bandwidth, facilitating communication between separate server nodes. While the UALink-over-Ethernet provides up to 260 TB/s aggregated scale-up bandwidth, aligning with Nvidia's NVL72 VR200 on paper, the underlying Ethernet transport remains a point of consideration for system architects.

The MI455X accelerators are expected to be among the best hardware available this year. However, Tom's Hardware notes that Helios systems with UALink interconnects over copper may have a limited market lifespan. AMD plans to launch its Instinct MI500-series products next year, which will likely be incorporated into next-generation rack-scale offerings, potentially requiring optical interconnects with UALink. This suggests Helios systems with true UALink over copper might be on the market for less than a year.

System architects evaluating Helios must weigh the advantages of an open, Ethernet-based fabric—namely cost and compatibility—against the potential performance degradation from higher latency and overhead. The choice between UALink-over-Ethernet and future 'true' UALink implementations will depend on workload requirements and the evolving maturity of the UALink ecosystem.

Sources

  1. 01 AMD's Helios MI455X AI platform breaks cover, initial systems use UALink-over-Ethernet interconnects — AMD's Vera Rubin rival surfaces, but the downsides of Ethernet could hamstring performance — Tom's Hardware