AMD Helios MI455X Brings a 72-GPU Rack Option to AI Data Centers

AMD Helios MI455X is not just another accelerator announcement. As of July 3, 2026, it is AMD’s clearest move toward selling an AI infrastructure blueprint at the rack level, rather than only competing GPU by GPU. Supermicro showed a 72-GPU double-width Helios rack at Computex 2026, built around AMD Instinct MI455X accelerators, 6th Gen AMD EPYC CPUs, AMD Pensando networking and the ROCm software stack. That matters because large AI deployments are increasingly judged as complete systems: compute trays, memory capacity, switching, cooling, software and serviceability all have to line up before a model-training cluster becomes useful. (ir.supermicro.com)

The headline configuration is a rack-scale design with 72 AMD Instinct MI455X GPUs paired with next-generation EPYC “Venice” CPUs. AMD’s published Helios and MI450-series figures point to up to 432GB of HBM4 memory and 19.6TB/s of memory bandwidth per GPU, with the full 72-GPU rack reaching 31TB of HBM4 memory and 1.4PB/s of aggregate memory bandwidth. AMD also lists up to 1.4 exaFLOPS of FP8 and 2.9 exaFLOPS of FP4 peak theoretical performance for the 72-GPU Helios rack, so the pitch is clearly aimed at massive model training and high-throughput inference rather than conventional enterprise acceleration. (amd.com)

View NVIDIA DGX Spark Personal AI Desktop Supercomputer on partner website

Helios is also where the conversation gets more technical. AMD’s platform is tied to open infrastructure ideas such as OCP Open Rack Wide, UALink and Ultra Ethernet, but early systems shown by partners have been reported with UALink-over-Ethernet scale-up connectivity. HPE’s Helios collaboration also describes a scale-up Ethernet switch, developed with Broadcom, using the Ultra Accelerator Link over Ethernet standard. The practical benefit is that Ethernet components, cables and operational knowledge are already familiar to cloud builders. The trade-off is that Ethernet-based scale-up fabrics may bring more latency, overhead or variability than a dedicated GPU fabric, which is a key detail for workloads that need all 72 GPUs moving together efficiently. (d1io3yog0oux5.cloudfront.net)

See SANDISK Optimus GX PRO 8100 PCIe 5 SSD price

The comparison to Nvidia is unavoidable because Nvidia’s rack-scale playbook has already shaped how buyers think about AI factories. Nvidia’s GB200 NVL72, for example, connects 36 Grace CPUs and 72 Blackwell GPUs in a liquid-cooled rack-scale design, while the newer Vera Rubin NVL72 platform is also framed around 72 GPUs connected through NVLink. AMD’s alternative angle is openness: Helios leans on ROCm, OCP alignment, Ethernet-based scale-out, Pensando networking and partner-built systems from companies such as Supermicro and HPE. That does not automatically make it faster or easier to deploy, but it gives hyperscalers and large enterprises another path if they want rack-scale AI hardware with a more standards-led design. (nvidia.com)

Buy ASUS GeForce RTX 5090 here

Helios should be viewed as an infrastructure platform still moving into the market, not a finished consumer-style product with one fixed spec sheet. HPE has said it will offer AMD Helios worldwide in 2026, while Supermicro positioned its Computex system for large-scale AI training and inference deployments. The open questions are the ones data center teams already care about: software maturity in ROCm, real cluster utilization, cooling density, power delivery, service procedures and how UALink-over-Ethernet behaves under production-scale training jobs. In other words, the MI455X rack is interesting because of the whole system around the GPUs. If AMD and its partners can make that system easy enough to deploy and predictable enough to operate, Helios becomes a serious 72-GPU alternative in a market that has been heavily shaped by Nvidia’s integrated rack strategy. (d1io3yog0oux5.cloudfront.net)

See AMD Radeon RX 9070 16GB price