GPUs aren't worth their weight in gold

"This being the HPC crowd, high-precision FP64 floating point computing is still an important aspect of the systems that run all manner of simulations and models. While there are tricks to take the solvers written for 64-bit processing and adapt them to the lower precision vector and tensor units that have been embedded in GPUs to accelerate AI training and inference, this is still not common practice and in fact we do not know how practical this approach is at scale."

"All things being equal, HPC centers would rather not have to gut their applications to get a 10x improvement in applications. Then again, they seem to be on a Sisyphean task of doing just that over the more than six decades since supercomputing as we know it - arguably starting with the CDC 6000 designed by Seymor Cray for Control Data Corp - came into being. The reason is that absolute performance often requires jarring architecture changes, ones that the mainstream of enterprise computing cannot stomach but which HPC centers exist precisely to bear."

November has long marked the supercomputing calendar, with SC25 in St. Louis drawing over 16,500 attendees and 559 exhibitors focused on HPC hardware and infrastructure. Many exhibitors emphasized cooling and GPU support for accelerated systems. High-precision FP64 floating-point capability continues to be crucial for simulations and models, while translating 64-bit solvers to lower-precision GPU tensor and vector units is possible but not yet widely adopted or validated at scale. HPC centers resist wholesale application rewrites despite potential speedups, reflecting a decades-long pattern in which pursuit of absolute performance demands disruptive architectural changes that enterprise computing often rejects.

Read at Theregister