When Richard Feynman and postdocs Stanley Frankel and Eldred Nelson received delivery of a computer at Los Alamos Laboratory in 1944, rather than wait for the IBM maintenance man to arrive to set the machine up, they had a go at it themselves, succeeding “after a fashion, but not satisfactorily”.1 In the years since, physicists have embraced computing as a way of seeing what goes on in systems too complex to solve with pen and paper or too inaccessible to experimental probes, and high-performance computing (HPC) and computational physics have co-evolved2 .But making the most of HPC requires recruiting and supporting a workforce of people who understand deeply both computers and the research questions they are used for: research software engineers.

There is much potential for specialists to unlock. Computational power has grown to an astonishing extent, with the exascale (1018 floating point operations per second, known as flops) era at hand. There are hopes that such power will enable projects like digital twins of fusion reactors — high-fidelity integrated models of not only the plasma but also the engineering aspects.

But there’s a big difference between achieving exaflop speeds on carefully controlled benchmarks and doing the same in the real world. A lot of physics simulations are run using older code, designed for earlier generations of computers. These software packages can’t take advantage of the latest hardware, and the users don’t always have the programming chops to make the necessary updates, which may involve fundamental changes to the algorithms.

The difference between old and new hardware can be dramatic. For instance, a major development in computing in recent years has been the rise of graphics processing units (GPUs), which originated in video game systems but are now widespread in HPC. GPUs are specialized for low-precision matrix multiplication at high speeds — ideal for computer graphics — whereas CPUs are generalized and can work at high precision. These differences determine what kinds of problems the hardware can be used to solve efficiently.

Furthermore, although physics and computing have a shared history, the future is being shaped by other computing applications. The adoption of GPUs has been driven in part by fields such as AI that are far more lucrative than scientific computing. These market forces are so strong that some analysts are raising the alarm about the risk that “future commodity hardware will not be appropriate for traditional modeling and simulation applications”3. In the face of these challenges, it’s essential that computational physics has access to the know-how to make the most of what’s out there and not be left behind.

Although physics’ need for computer expertise is ever more pressing, there have always been productive collaborations. Two days after Feynman, Frankel and Nelson received their computer, one of the best in IBM’s maintenance team, John Johnston arrived and adjusted the machine to get it into proper working order. Today, some universities have research software groups, which employ research software engineers who can support the work of other research groups across the sciences.

Such collaborations exist in part thanks to organizations like the Society of Research Software Engineers in the UK, which was established in 2013 out of a grass-roots discussion among software specialists working in science;4 similar movements have since been established elsewhere in the world. The society supports research software engineers and campaigns about issues like the need for academic career paths for software professionals, who may be employed on back-to-back postdoc contracts for many years, owing to a lack of suitable permanent positions.

There is also a need to attract people into research software engineering in the first place. One thing that can help is strong communication between researchers with a physics focus and those with a software focus, because many software specialists come from research backgrounds. At Nature Reviews Physics, we want to foster this dialogue, so in this issue, we publish three pieces on what computational research in fusion and quantum mechanical simulations may look like in the coming years (see the Perspective by Marta Garcia-Gasulla and Mervi J. Mantsinen, the Perspective by Frank Jenko and the Perspective by Ravindra Shinde and colleagues). We hope these pieces provoke further conversation.