AMD’s memory advantage targets growing demand from AI agents
AI agents require substantial high-bandwidth memory to retain working context while using tools and completing tasks. That demand favors AMD’s emphasis on memory capacity: its MI455X carries 432 gigabytes of HBM4, compared with 288 gigabytes on Nvidia’s competing Rubin chip. More capacity can accommodate additional active agents without moving their context into slower memory.
AMD also lists memory bandwidth of 23.3 TB/s, versus Nvidia’s revised 19.2 TB/s for Rubin. These are specifications, not real-world benchmarks. Damnang Research estimates that a Helios rack could support about 66% more simultaneous agents than a Rubin rack, assuming the same model and a 150,000-token context per agent. Additional agents must share bandwidth, so capacity gains do not translate directly into equivalent throughput gains.
AMD claims Helios offers up to 30% more tokens per dollar than Nvidia’s Vera Rubin rack. Lisa Su has also cited 10% to 50% more AI output at the same rack power. Those projections rely on AMD’s lab modeling rather than head-to-head tests on shipping hardware.
The design carries trade-offs. AMD needs more memory stacks per chip during an HBM shortage, increasing supply exposure. Nvidia is also expanding memory capabilities and offers a context-storage platform for agents waiting on tools. AMD’s earlier memory lead did not secure substantial market share, underscoring the importance of software compatibility alongside hardware specifications.