News
AMD AI NICs Enable HPC and AI Integration for Scientific Advancements
AMD expands AI NIC technology to support HPC workloads, enabling AI and scientific simulations on a unified Ethernet network.
AMD has announced an expansion of its AI NIC technology to support high-performance computing (HPC) workloads, enabling a unified Ethernet network for large-scale AI training and MPI-based scientific simulations. This advancement is central to the operations of Lux, the AMD-powered AI factory for science at Oak Ridge National Laboratory (ORNL), ranked No. 14 on the November 2026 TOP500 list of supercomputers.
Unified Networking for AI and Scientific Computing
Lux integrates over 4,000 AMD Instinct™ MI355X GPUs, AMD EPYC™ Server CPUs, and AMD Pensando™ Pollara 400 AI NICs. Each AI NIC provides 400 Gb/s of scale-out connectivity, allowing Lux to run both AI and HPC workloads on a single Ethernet fabric. This shared network supports the U.S. Department of Energy’s Genesis Mission, which focuses on projects such as AI-assisted fusion power plant design, materials development, and Earth system prediction.
AMD AI NIC technology is designed to meet the distinct demands of AI and HPC workflows. AI training involves moving large volumes of data through collective operations, while scientific simulations rely on frequent exchanges and synchronization across processors. The programmable Ethernet foundation of AMD AI NICs supports both types of workloads and can evolve through software and firmware updates.
Performance Benchmarks and Scalability
On the AI front, training a 123-billion-parameter language model using AMD Instinct™ MI355X GPUs and AMD Pensando™ Pollara 400 AI NICs achieved 4.48 exaflops of FP8 hybrid-precision throughput on 2,048 GPUs. Scaling the job 16-fold from 128 GPUs retained 98.3% of per-GPU throughput, with GPUs sustaining over 85% of their peak BF16 compute rate.
For scientific simulations, AMD AI NIC technology demonstrated robust performance:
- Collective latency: An 8-byte MPI Allreduce operation across 2,048 GPUs completed in 103 microseconds on average, with latency increasing by only 17% when scaled to 3,760 GPUs.
- Molecular dynamics: Using the LAMMPS code with the SNAP machine-learning interatomic potential, a simulation of 17.2 billion tungsten atoms ran 40.9 times faster on 3,200 GPUs compared to 64 GPUs, maintaining strong-scaling efficiency above 81.9% at the largest scale.
- Turbulence simulations: In weak-scaling runs of the M-PSDNS spectral turbulence solver, the system processed approximately 188 billion grid points per second on 1,024 GPUs, achieving 72.7% FP32 weak-scaling efficiency relative to a single node.
Implications for Research and Development
AMD AI NIC technology enables research organizations to use a common Ethernet foundation for both AI and HPC workloads, facilitating the integration of AI models and scientific simulations into unified workflows. This approach supports evolving standards and workloads, ensuring long-term adaptability for supercomputing systems.