AMD has introduced Helios, a rack-scale AI platform that combines Instinct MI455X GPUs, EPYC Venice CPUs, Pensando networking and the ROCm software stack. Its use of UALink over Ethernet is intended to give data-center operators a more open alternative to tightly integrated proprietary systems.
A complete rack design
Helios combines compute, memory and networking at rack scale so that data movement is considered alongside GPU throughput. AMD cites 432GB of HBM4 memory per MI455X, 31TB across a rack and up to 1.4PB/s of bandwidth.
Open connectivity as a differentiator
NVIDIA’s Vera Rubin NVL72 is optimized around NVLink, while Helios is built around UALink over Ethernet, developed with industry partners. A shared interconnect could make it easier to combine accelerators, CPUs and networking equipment from multiple vendors. Real-world results will depend on software and system integration.
Shipping target
AMD says hardware partners are expected to begin shipments by the third quarter of 2026. Production will test GPU supply, networking, cooling and rack integration together.
Takeaway
Helios shows AMD moving from accelerator products toward a full rack-scale AI proposition. Its competitive case will rest on open connectivity, software usability and multi-vendor deployment as well as peak performance.
Source: AMD official announcement (accessed August 10, 2026).
