- A provocative paper questions the need for GPUs, arguing that CPUs are evolving with vector and matrix engines, HBM memory, and integrated interconnects.
- The discussion is prompted by the all-CPU 'LineShine' supercomputer, the world's fastest AI/HPC system in recent Top500 rankings.
- Historically, GPUs gained prominence due to CPUs lacking sufficient flops and memory bandwidth, but CPU designs are now catching up with enhanced math capabilities.
- Examples include Arm's SVE/SME and Intel's AMX units, with IBM having integrated matrix engines earlier in Power and mainframe processors.
- Hybrid GPU systems require code splitting and data movement, adding cost and complexity, whereas integrated CPU designs could streamline AI and HPC convergence.
- The paper suggests that CPUs with advanced features can eliminate the fundamental need for discrete GPUs, especially for combined simulation and AI workflows.
- Process node advancements (e.g., moving to 3nm) could significantly boost the efficiency of CPU-only systems like LineShine, rivaling GPU-accelerated machines.
- Future trends may see Nvidia integrating tensor cores into Arm server CPUs, creating unified architectures that could run CUDA software without offload models.