AMD MI430X targets 288 TFLOPs of FP64 as AI reshapes HPC
AMD’s MI430X targets 288 TFLOPs of FP64 compute as Oak Ridge links AMD systems with IBM’s 129-qubit Nighthawk for DOE-backed quantum error-correction research.
- By
- Rebecca Stone
- Filed
- Channel
- AI & Compute
- Read
- 3 min read
By Rebecca Stone
AMD’s Instinct MI355X, then one year old, peaks at 78.6 TFLOPs of FP64 compute, while the HPC-oriented Instinct MI430X is expected to peak at 288 TFLOPs.
Nvidia’s latest Rubin GPU has 33 TFLOPs of FP64 compute, below the 34 TFLOPs of the four-year-old H100. Across Nvidia’s last two GPU generations, native double-precision performance has appeared to stall as the company prioritized lower-precision formats for AI; Nvidia has leaned on emulation to lift FP64 throughput.
The gap matters because FP64 supplies the precision required by high-accuracy simulations, while AI workloads often use approximate, lower-precision formats.
Scott Tease, AMD’s corporate vice president for HPC and sovereign AI, said ahead of SC26 that most HPC sites “recognize that AI will never replace HPC.” He expects high-precision workloads to “exist for a very long time.” Tease joined AMD earlier this year after more than 10 years at Lenovo and 13 at IBM.
“Maybe a few years ago, people in HPC might have felt a little threatened by AI because it was taking some of the limelight,” Tease said. He added that researchers now see AI as essential to scientific discovery. The harder question is deployment.
What does a heterogeneous stack solve?
Tease said researchers now spawn dozens of AI agents to explore possible paths toward a given intent, while complex simulations stay on traditional FP64 systems. CPUs and GPUs arranged as disconnected “islands” cannot sustain those connected workflows, he said.
AMD, IBM and Oak Ridge National Laboratory are testing a more integrated design through the Umbra HPC system at the Oak Ridge Leadership Computing Facility. It places AMD’s classical computing stack alongside IBM’s 129-qubit Quantum Nighthawk quantum computer and connects both to the facility.
The stated aim is to explore how quantum, AI and scientific computing can accelerate discovery and national-security work. The project also provides a shared environment for evaluating the architectures, software and workflows required by future hybrid quantum systems.
IBM describes the Oak Ridge effort as a co-design initiative that brings quantum and classical computing into one system. For researchers, that means testing quantum error correction with classical systems, including ways to suppress environmental effects that corrupt quantum states.
When should a workload move to quantum?
Open-source Slurm already schedules much of the TOP500, so basic placement of classical HPC jobs is largely established. The Oak Ridge effort targets a harder orchestration layer: deciding not only where a workload should run, but when GPU acceleration adds value and when quantum is appropriate.
The U.S. Department of Energy’s Quantum Genesis Q competition will provide up to $215 million for quantum research. By late 2028, it calls for systems with at least 100 logical qubits capable of hundreds of millions of fault-tolerant operations.
Those remain program targets. The 129-qubit Nighthawk is the machine IBM is deploying with Oak Ridge, and IBM calls the error-correction work there a key component in reaching the competition goal.
Tease expects scientific computing to converge on a unified fleet in which AI, FP64 HPC and quantum coexist. He does not argue that quantum solves everything. In human health and chemistry, he expects quantum to tackle work classical computers cannot do today; elsewhere, classical resources will remain central.
What changes beyond Oak Ridge?
Separately, AMD announced the previous week that it had invested in quantum company PsiQuantum. PsiQuantum will use AMD CPUs, GPUs and FPGAs for its first fault-tolerant quantum system, extending the chip supplier’s role beyond conventional HPC.
The next competitive test is whether AMD can deliver the MI430X’s 288-TFLOP FP64 target while the Oak Ridge pilot turns AI, FP64 and quantum scheduling from a shared test bed into a repeatable workflow.
Source: Tom's Hardware
More from Rebecca Stone
Show full bio
Correspondent covering technology and product news in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,037 articles
Related articles
amd-buys-taalas-to-hard-wire-ai-inference-into-silicon-5aa48930
AMD Buys Taalas to Hard-Wire AI Inference Into Silicon
nvidia-targets-availability-for-rubin-hpc-science-gpu
Nvidia targets 2028 availability for Rubin HPC science GPU
intel-d270-claims-5-9x-fp64-throughput-lead-over-nvidia-rtx-pro-6000
Intel D270 claims 5.9x FP64 throughput lead over Nvidia RTX PRO 6000
microsoft-commits-billions-to-amd-helios-racks-for-azure-inference-1f4c04ad
Microsoft Commits Billions To AMD Helios Racks For Azure Inference


