The computing hardware of today is vastly different to what was available when
the LHC experiments began writing their software, almost 20 years ago. Single
CPU cores with ever faster clocks have given way to multi-core chips with wide
vector registers for parallel processing. At the same time, Graphics Processing
Units (GPUs) have become ever more powerful and ever more popular, offering
thousands of cores and raw floating point performance much greater than CPUs can
manage. This presents two huge challenges to HEP software. The first is
concurrency, the adaption to performing many tasks in parallel during data
processing. The second is heterogeneity, where instead of a mono-culture of
x86_64 processors, HEP needs to adapt to a mixture of different CPU
architectures, GPUs and even more exotic processors, such as FPGAs. These
challenges from modern hardware come just at the time when planning for the
High-Luminosity LHC foresees a x10 increase in rate for ATLAS and CMS and a
tremendous jump in event complexity, with pile-up perhaps reaching 200.
To face these challenges computing and software experts from multiple
experiments and institutions came together in 2015 to begin a bottom-up
organisation to tackle the challenges of the coming years and to avoid each
experiment having to work alone to solve these problems. The HEP Software
Foundation (HSF) undertook to organise the community to write a Community White
Paper Roadmap, where more than 300 physicists and computing experts came
together to map out a plan for progress from event generation through to final
analysis. Since the CWP publication there has been huge progress in HEP to
tackle these problems, though much remains to be done. I will touch on the most
important developments that have been, and remain to be, done. Finally, I will
end on the question of how we argue for more resources to fund software in the
future and best work together as a community with other data intensive sciences
and industry.