During Hot Chips 2025, AMD provided a comprehensive overview of its latest Instinct MI350 series, with a particular focus on the high-density MI355X DLC (Direct Liquid Cooled) rack system. The MI355X DLC rack represents a significant leap in AI and HPC infrastructure, offering scalability from individual nodes up to large rack deployments.
The MI355X DLC rack is a 2OU (Open Compute Project Universal) design that integrates 128 MI355X GPUs and 36TB of HBM3e high-bandwidth memory. At FP4 precision, the system achieves a peak performance of 2.6 exaflops, supporting highly demanding AI and data workloads. AMD also offers a 96-GPU EIA (Enterprise Integration Architecture) version with 27TB of HBM3e.
The underlying platform features 5th Gen AMD Epyc CPUs, MI350 series GPUs, AMD Pollara network interface cards, and OCP-standard form factors, all interconnected via Infinity Fabric with a throughput of up to 1,075GB/s. Bandwidth, memory capacity, and compute density have been significantly enhanced compared to previous generation installations.
AMD’s node-level configurations support both air and direct liquid cooling. For reference, an MI350X air-cooled system with eight GPUs achieves 73.8 petaflops at FP8, while the denser liquid-cooled MI355X system reaches 80.5 petaflops at the same precision.
Looking to future developments, AMD’s roadmap details recent and upcoming launches: the MI325X debuted in 2024, the current MI350 family arrived in 2025, and the Instinct MI400 is expected in 2026, offering up to 40 petaflops FP4, 432GB HBM4 memory, and advanced data throughput figures.
In the broader market context, Nvidia’s Vera Rubin platform is scheduled for commercial release in 2026-27. The Vera Rubin NVL144 system, based on Nvidia’s future architecture, will be rated for 3.6 exaflops FP4 inference and 1.2 exaflops FP8 training, supported by 13TB/s HBM4 bandwidth and 75TB of high-speed memory. The roadmap further projects the Rubin Ultra NVL576 for release in 2027, scaling up to 15 exaflops FP4 and 365TB total memory, with extremely high interconnect bandwidth via NVLink7 and CX9 standards.
AMD’s MI355X DLC rack, as described, represents an available production solution for 2025, whereas Nvidia’s Vera Rubin systems remain in the pre-release or roadmap phase for subsequent years. The comparative data provides insight into each manufacturer’s strategy for high-performance AI infrastructure, with AMD offering an attainable platform in the near term and Nvidia positioning for longer-term, larger-scale deployments.
