AMD Helios rack-scale system delivers 15% more AI compute density than Nvidia Vera Rubin NVL72, matching performance on standard workloads.

AMD has officially launched Helios, a rack-scale supercomputer solution designed for AI infrastructure and large-scale foundational-model training and inference. The system combines 72 AMD Instinct MI455X GPUs with the latest 6th generation AMD EPYC 'Venice' server CPUs, integrated with AMD ROCm software and AMD Pensando networking to deliver 2.9 ExaFlops of FP4 compute performance.

In direct comparison with NVIDIA's Vera Rubin NVL72 rack-scale solution, AMD claims Helios delivers 15% more peak FP4 compute performance. The memory advantage is even more pronounced: Helios features 31TB of HBM4 memory capacity, representing 50% more than Vera Rubin's offering. The system provides 43 TB/s scale-out bandwidth and 1.7 PB/s of memory bandwidth.

According to AMD, these specifications translate to tangible economic benefits for customers deploying large-scale AI infrastructure. The company claims Helios can deliver 30% more tokens per dollar compared to competing solutions, an important metric for AI companies looking to optimize their infrastructure costs as demand for AI training and inference capacity continues to grow.

AMD positions Helios as a building block that enables customers to scale efficiently from a single rack to large gigawatt-scale AI clusters while maximizing performance, performance per watt, and cost per token. The integrated rack-scale design is intended to simplify deployment while providing the compute, networking, and open software foundation required for next-generation AI infrastructure deployments.