Nvidia is proposing a beast of a CPU system for Windows PCs
Points and comments are a snapshot, not live.
Nvidia proposes a Windows PC CPU with 128 GB unified memory and up to 6,144 CUDA cores for local AI inference.
Nvidia's new chip for Windows PCs features 10 performance cores and 10 efficiency cores based on Cortex-X925, paired with up to 6,144 CUDA cores. The standout feature is 128 GB of shared unified memory, eliminating the separate CPU and GPU memory pools that traditionally limit local AI model deployment. This matches the approach Apple pioneered years ago. The unified memory is slower than dedicated GPU memory but provides sufficient bandwidth for running AI models locally while remaining cost-effective. Lemire notes the chip supports six 128-bit SIMD execution units (SVE2), inferior to AMD's AVX-512 but better than Apple Silicon on paper. Primary use cases include gaming and local AI inference, though Lemire considers local AI models still niche.
What commenters are saying
Commenters split on viability. One camp highlights memory bandwidth limitations: a user with a 128 GB LPDDR5X machine reports bandwidth is "awful" for AI, and old server CPUs outperform it. Another camp notes AMD's competitor, Strix Halo, already offers 192 GB unified memory with better CPU multi-threading performance via AVX-512, making Nvidia's sole advantage a larger GPU (6,144 vs 2,560 cores) at higher cost. A commenter clarifies the chip already ships in DGX Spark systems with real benchmarks available, not a proposal. Discussion centers on whether local model economics justify the hardware cost given remote API pricing uncertainty and quantized model quality.