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
AMD’s first rack-scale AI machine is incoming.
Several AMD partners are showing off the company’s next-generation Helios rack-scale solution running AMD’s EPYC ‘Venice’ processors and Instinct MI455X AI accelerators at Computex 2026 in Taipei, Taiwan. The units are set to become available later this year. There is one major catch, though: they all use UALink-over-Ethernet scale-up connectivity, which may limit their performance in certain workloads that depend on the connection performance. That said, Helios systems with ‘true’ UALink interconnects will also be available.
AMD’s Helios is the company's first rack-scale AI system, and is set to rival Nvidia’s NVL72 VR200 machines based on the next-generation Vera Rubin platform. Helios will rely on AMD’s 6th Generation EPYC Venice CPUs with up to 256 cores, pack 72 Instinct MI455X accelerators with a total of 31 TB of HBM4 memory, and 1400 TB/s of bandwidth. AMD estimates that its performance will be around 2900 FP4 dense PFLOPS, which puts the unit behind Nvidia's VR200 NVL72 system in terms of compute performance, but ahead of it with HBM4 memory capacity. This promises to provide Helios-based systems an advantage in memory-intensive workloads, such as when running large LLMs.
The AI accelerators are interconnected and make use of a UALink-over-Ethernet connection, which provides up to 260 TB/s aggregated scale-up bandwidth (in line with Nvidia’s NVL72 VR200). Helios will also feature Pensando Vulcano network interface cards (NICs), which are among the industry's first 800 GbE network cards that comply with the Ultra Ethernet specification and provide up to 43 TB/s of scale-out bandwidth.
However, the interconnection used on these Helios systems will vary. The machine supports both UALink and UALink-over-Ethernet, but the initial versions will use the latter, rather than the former. This is likely because UALink switches aren't finalized and are pending validation and qualification by AMD’s AI customers.
The biggest advantage of using UALink over Ethernet is that AMD can build Helios using an existing, widely supported ecosystem of validated and qualified components. Ethernet switching ASICs, cables, and other ingredients are already used by hyperscalers and cloud providers worldwide, which accelerates deployment.
But there is a major downside with using Ethernet, even with the UALink protocol on top: Ethernet was originally designed as a general-purpose networking technology; it was never designed to scale up AI accelerators.
As a result, communications may involve higher latency, more protocol overhead, and less deterministic performance than a dedicated scale-up fabric. For large AI training jobs that need all 72 Instinct MI455X accelerators to work in concert, communication efficiency is as important as compute performance. If the UALink-over-Ethernet interconnect cannot keep GPUs fed with data efficiently, some of the theoretical performance of the hardware may be lost in real-world deployments, even though on paper, Helios with UALink-over-Ethernet is as good as Nvidia’s NVL72 VR200 in scale-up bandwidth.
This begs the question of whether UALink will ever be widely used with Helios and whether UALink will ever be widely deployed using copper. Hyperscalers and other companies deploying high-end AI hardware at scale rarely upgrade their hardware.
While the Instinct MI455X certainly promises to be among the best hardware accelerators this year, Helios will likely only be rivalled by Nvidia’s NVL72 VR200. It will be outdated next year when AMD launches its Instinct MI500-series products. These units will be used in the company’s next-generation rack-scale offering, which promises to pack more AI GPUs, potentially requiring optical interconnects with UALink on top. As a result, Helios systems with true UALink interconnections over copper will be on the market for less than a year before those next-generation rack-scale solutions will hit the market.
Of course, nothing is stopping AMD from offering Helios with Instinct MI500-series accelerators and UALink interconnects over copper; however, the company hasn't confirmed the existence of such systems.
Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.