AMD denies report of MI455X delays as Nvidia VR200 systems are rumored to arrive early — company says Helios systems 'on target for 2H 2026'
Summary
AMD's Helios rack-scale solution, powered by MI455X accelerators, faces a potential delay until 2027, while Nvidia is rumored to expedite the launch of its Vera Rubin platform, intensifying competition in the AI market.
Key Insights
What is AMD Helios and why is it significant in the AI market?
AMD Helios is a rack-scale system solution for high-performance computing deployments that combines 72 Instinct MI455X accelerators with 31 TB of HBM4 memory in total. The system is designed to deliver up to 2.9 exaFLOPS of FP4 compute performance for AI inference and 1.4 exaFLOPS for AI training. Helios represents AMD's direct competition with NVIDIA's Vera Rubin platform in the enterprise AI infrastructure market, where data centers require massive computational power for training and deploying large language models and other AI applications.
How does the MI455X compare to NVIDIA's Rubin accelerator in terms of memory capacity?
The MI455X features 432GB of HBM4 memory per accelerator, which exceeds NVIDIA Rubin's 288GB capacity. This 2.25x increase in memory headroom provides AMD's clearest competitive advantage, enabling inference on larger AI models without requiring tensor parallelism—a technique that splits models across multiple GPUs. This reduces system complexity and interconnect demands, making it easier to deploy very large language models on AMD's infrastructure.