Nvidia Proposes High-Performance CPU System for Windows PCs

Targeting demand for running local AI models, Nvidia has proposed a new Windows PC system architecture in which the CPU and GPU share 128GB of shared memory.

This system consists of up to 6,144 CUDA cores and a combination of 10 performance cores and 10 efficiency cores based on Cortex-X925. The key differentiator is a Unified Memory structure similar to Apple Silicon, where the CPU and GPU share a single memory pool, aiming for zero-copy operations that reduce data transfer overhead.

Key features and considerations are as follows:

• AI Optimization: While bandwidth may be lower than dedicated GPU memory, it provides the memory capacity and cost efficiency needed to run large-scale local AI models.
• Performance Comparison: CPU performance is expected to be superior to Apple Silicon, but there may be differences in versatility compared to AMD's AVX-512 support.
• Constraints: Since the memory is soldered onto the board, post-purchase upgrades are not possible, and memory bandwidth could become a bottleneck for actual performance.

This proposal focuses on meeting the needs of prosumers who want to run local LLMs and enterprises that prioritize security.