Three HPC Gurus Ask: Do We Still Need GPUs? (www.nextplatform.com)

In a thought-provoking new paper, HPC experts Jack Dongarra, Torsten Hoefler, and Satoshi Matsuoka pose the question: "Do We Still Need GPUs?" Their inquiry comes as the rapidly evolving CPU landscape, particularly with hybrid vector and matrix math engines, challenges the traditional dominance of GPUs in AI and high-performance computing (HPC). The authors highlight the recent launch of LineShine, the world’s fastest all-CPU supercomputer, which has demonstrated that CPUs can now effectively handle the demands of AI workloads alongside traditional simulations, contrasting with the GPU-centric approaches of the past. The significance of this discussion lies in the potential for CPUs to integrate the capabilities that once necessitated GPUs, such as advanced vector and matrix operations along with high-bandwidth memory. This shift could facilitate the convergence of AI and computational science, enabling workflows that combine modeling, AI, and data analysis efficiently within a single architecture. For the AI/ML community, this evolution suggests that future systems could rely less on GPU hardware, reducing costs and energy consumption while maintaining high performance. With improvements in CPU capabilities and manufacturing processes, we may be witnessing a pivotal moment in computing that redefines hardware requirements for AI and complex simulations.