Skip to content
Theme
OnDevice Tools

GadgetsLaptops & PCs

AI will never replace HPC, says AMD executive as FP64 demand endures

AMD's Scott Tease says high-precision computing remains essential to science even as AI expands, with hybrid systems combining AI, HPC and quantum hardware emerging as the way forward.

Illustration of a supercomputer data centre combining AI accelerators and classical processors

Key takeaways

  • AMD's Scott Tease says HPC sites recognise AI will never replace high-precision supercomputing.
  • Nvidia's Rubin GPU offers 33 TFLOPs of FP64 compute, below the H100's 34 TFLOPs, while AMD's MI430X targets 288 TFLOPs.
  • AMD and IBM are deploying a hybrid system at Oak Ridge combining the Umbra HPC machine with a 129-qubit quantum computer.
  • The US Department of Energy's Quantum Genesis Q competition offers up to $215 million and targets 100 logical qubits by late 2028.
In this article
  1. What AMD says about AI and HPC
  2. Why FP64 precision matters
  3. A hybrid compute stack at Oak Ridge
  4. Quantum error correction and funding
  5. What it means for researchers

What AMD says about AI and HPC

AMD does not believe artificial intelligence will displace traditional supercomputing, according to Tom's Hardware. Scott Tease, the company's corporate vice president of HPC and sovereign AI, told the publication ahead of the SC26 conference that most high-performance computing sites recognise AI will never replace HPC.

Tease said researchers once felt threatened by AI because it was taking the limelight, but they now see it as an essential part of scientific discovery. The question, in his view, is how to deploy it alongside existing simulation workloads rather than instead of them.

Why FP64 precision matters

High-performance computing relies on FP64, a double-precision number format that allows extremely accurate calculations for simulations in science and engineering. AI workloads, by contrast, use low-precision data formats that trade accuracy for speed.

Tom's Hardware notes that Nvidia's FP64 performance has stalled in recent generations as the company focused on AI. Its Rubin GPU offers 33 TFLOPs of FP64 compute, below the 34 TFLOPs of the four-year-old H100. AMD's Instinct MI355X peaks at 78.6 TFLOPs, while its upcoming MI430X reaches 288 TFLOPs. Nvidia has leaned on emulation to raise FP64 performance, a tactic AMD is pushing back on.

A hybrid compute stack at Oak Ridge

Tease argues the best path forward is a heterogeneous compute stack that brings AI, HPC and quantum hardware together. AMD is helping pilot this approach at Oak Ridge National Laboratory alongside IBM.

The systems will connect to the Oak Ridge Leadership Computing Facility, where researchers will explore how integrated quantum, AI and scientific computing can accelerate discovery and national security work.

Tease said islands of CPUs or GPUs that are not interconnected will not keep up. He envisions a single fleet where software decides when a problem suits AI, GPU acceleration or a quantum environment, going beyond traditional schedulers such as Slurm.

Quantum error correction and funding

Illustration: science, business, AI, AMD logo and the labels AMD, HPC, IBM

One goal of the Oak Ridge system is accelerating quantum error correction, the process of weeding out environmental influences that disturb quantum calculations, using classical hardware. IBM describes this research as a key component of its quantum roadmap.

The work supports the Quantum Genesis Q competition launched by the US Department of Energy, which will provide up to $215 million in quantum research funding.

Separately, AMD announced last week an investment in PsiQuantum to supply CPUs, GPUs and FPGAs for the company's first fault-tolerant quantum system.

What it means for researchers

The message from AMD is that scientific computing's future is coexistence rather than replacement. Tease said problems in fields such as human health and chemistry will need quantum systems, while complex simulations will continue to run on FP64-driven HPC fleets for a very long time.

For research institutions, the shift points towards unified infrastructure that can route each part of a workload to the most suitable hardware. Whether that vision arrives quickly depends on how well vendors can interconnect systems that today often operate in isolation.

Sources

  1. Researchers ‘recognize that AI will never replace HPC,’ says AMD executive — high-precision workloads will ‘exist for a very long time’ even as AI usage expandsTom's Hardware

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.