US partners with AMD on $1b supercomputer, AI projects

AMD and the US Department of Energy will provide two new supercomputers, Lux AI and Discovery, for Oak Ridge National Laboratory to advance AI and scientific research.

Lux AI, co-developed by ORNL, AMD, Oracle Cloud Infrastructure, and HPE, will be deployed in early 2026 and focus on AI-enabled science, energy, materials, medicine, and advanced manufacturing.

Discovery, expected at Oak Ridge in 2028, will use next-generation AMD EPYC CPUs (“Venice”) and AMD Instinct MI430X GPUs to support high-performance computing and AI.

Both systems are intended to support the US AI Action Plan and represent a combined investment of about US$1 billion from public and private funding.

Discovery will build on the breakthroughs of Frontier, aiming for improved performance, energy efficiency, and AI capabilities.

Once operational, Discovery is set to become a key part of US AI computing infrastructure from 2029.

🔗 Source: AMD

🧠 Food for thought

Implications, context, and why it matters.

Missing concrete specifications leave revenue and timeline uncertainty for vendors

Coverage cites $1 billion yet omits details that define impact. Discovery targets 3 to 5 times Frontier throughput on benchmarks 1 without node counts, peak floating point operations per second (FLOPS), or a power envelope.
Frontier at Oak Ridge National Laboratory uses 9,472 nodes with AMD Epyc 7713 CPUs and 37,888 Instinct MI250X GPUs 2. Discovery and Lux AI list no scale, and a jump from 10,000 to 50,000 nodes would shift spend by billions, stretching schedules.
The MI430X for Discovery still lacks public detail 3, and planning for Lux AI cites AMD Instinct MI355X 3 without memory capacity, interconnect bandwidth, or power use. Those gaps block judgments on whether these builds rival Frontier liquid cooling that hit about 62.86 gigaflops per watt 2.

CUDA to HIP migration needs rise as Oak Ridge National Laboratory (ORNL) next systems use AMD GPUs

Both builds standardize on AMD GPUs, so CUDA code holders must port to Heterogeneous-Compute Interface for Portability (HIP). Hipify-clang and hipify-perl automate about 90 to 99 percent 4, but the rest needs architecture specific work beyond tools 5.
IT consultancies and high performance computing specialists can package CUDA to HIP migration, tuning, and training. LAMMPS and HACC ports prove feasibility 4, while sparse docs for AMD Instinct MI430X plus MI355X data center GPUs create room for targeted optimization services 3.
Universities and national labs need practical HIP courses. Curricula should cover launch bounds optimization, unified address space differences 6, plus atomics based patterns before Discovery delivery in 2027 or early 2028 1.