AMD has unveiled the Instinct MI455X and its accompanying Helios rack scale architecture at its recent technology showcase. The hardware platform links 72 GPUs into a single coherent memory system to directly challenge Nvidia in the enterprise market. However, software optimization remains the main challenge for customers looking to extract the full theoretical performance of these massive systems.
The hardware represents a major step forward for the CDNA 5 architecture. Inside each MI455X silicon package sits 320,000,000,000 transistors built using advanced packaging. AMD stacks 4 Accelerator Complex Dies fabricated on TSMC 2N gate all around technology on top of its Fabric and Cache Dies. This chiplet layout allows the company to use premium processes where they help performance while using more cost effective N3P nodes for the remaining components.
The upgrade to CDNA 5 alters how the GPU processes instructions. The wavefront width drops to 32 to improve instruction latency and lower register pressure. Peak mathematical performance also jumps, with MXFP4 format calculations reaching 40.26 PF. To help move all this data efficiently, the new Tensor Data Mover shifts data directly from system memory to the local cache without clogging up intermediate cache levels. Multicast memory reads also help by sending the same weights to multiple compute elements simultaneously, saving power on redundant bus traffic.
Memory capacity is where AMD holds a clear advantage over Nvidia Rubin hardware. The MI455X moves to HBM4 memory, offering 432GB of capacity and 23.3 TB/s of bandwidth per GPU. By comparison, the initial Rubin offering tops out at 288GB. This memory disparity is even more visible at scale. A full Helios rack aggregates 31.1 TB of total memory space. This is 50% more memory capacity than Nvidia equivalent setup, making it highly attractive for running massive artificial intelligence models entirely in memory.
The Helios platform packages these processors alongside EPYC Venice CPUs and Pensando network hardware. A single rack delivers 4,600 CPU cores and 18,000 GPU compute units. Production is currently underway with shipments planned to ramp up through the end of the year and into early next year. Tech giants are already signing on, with Microsoft planning scale deployments on Azure and Anthropic securing a major power supply deal. Looking further ahead, AMD confirmed that the upcoming MI500 series remains on schedule for next year with the MI600 series planned for the following year.
