Three HPC Gurus Ask: Do We Still Need GPUs?
The Next Platform, Tuesday, June 30th, 2026
Leading HPC researchers argue that CPUs with advanced matrix engines may eliminate the need for discrete GPUs.
Three prominent HPC experts contend that GPUs are becoming unnecessary as CPUs evolve to incorporate accelerator-class features. Modern processors now include vector extensions (SVE/AVX), matrix units (SME/AMX), high-bandwidth memory, and multiple precision formats that historically made GPUs essential.
The authors highlight China's all-CPU LineShine supercomputer as evidence that trillion-parameter AI models can run efficiently on enhanced processors.
They argue that integrated matrix acceleration in CPUs offers cleaner platforms for converged AI and scientific computing, reducing the complexity of offloading data between discrete devices. While GPUs remain powerful, they are no longer fundamentally required when CPUs match their computational capabilities.