What is NVLink-C2C? Chip-to-Chip Interconnect Technology
Quick Summary
- Definition: Chip-to-chip interconnect for CPU-GPU, GPU-GPU, GPU-DPU
- Bandwidth: 900 GB/s, 7x faster than PCIe Gen5
- Coherence: Cache-coherent, shared memory between CPU and GPU
- Use Case: Grace Hopper superchip uses NVLink-C2C
- Benefit: Eliminates CPU-GPU data copy bottleneck
What is NVLink-C2C? Chip-to-Chip Interconnect Technology
NVLink-C2C is NVIDIA's chip-to-chip interconnect technology that Grace Blackwell with NVLink-C2C provides ultra-high-bandwidth, low-latency connectivity between different types of processors—CPU to GPU, GPU to GPU, and GPU to DPU (Data Processing Unit). Introduced with the Grace Hopper superchip, NVLink-C2C represents a fundamental advancement in heterogeneous computing architecture by enabling cache-coherent shared memory between CPU and GPU.
Technical Specifications
NVLink-C2C delivers 900 GB/s of bidirectional bandwidth between connected processors—7x faster than PCIe Gen5's 128 GB/s. The interconnect is cache-coherent, meaning the CPU and GPU share a unified memory view without explicit data copying. Latency is approximately 100-200 nanoseconds, compared to 1-3 microseconds for PCIe transfers.
Grace Hopper Implementation
The GH200 Grace Hopper superchip uses NVLink-C2C to connect an ARM-based Grace CPU with a Hopper H100 GPU. The unified memory pool of 624GB (480GB LPDDR5X + 144GB HBM3e) is accessible by both processors without PCIe involvement. Applications running on the CPU can directly access GPU memory and vice versa, eliminating the data copy bottleneck that limits performance in traditional CPU-GPU architectures.
Impact on AI Workloads
NVLink-C2C particularly benefits AI workloads with frequent CPU-GPU data exchange. Graph neural network training, where CPU sampling and GPU computation alternate frequently, shows 3-5x speedup on Grace Hopper versus traditional x86+H100 systems. Recommendation systems with large embedding tables benefit from the unified memory capacity. Database acceleration and vector search workloads achieve 4-8x throughput improvement.
Related Content
Explore more about this topic:
- GPU Memory Bandwidth: Complete Guide
- FP8 vs FP16 vs BF16 vs FP32: Precision Formats
- Enterprise GPU Memory Hierarchy
Is NVLink-C2C the same as NVLink?
NVLink-C2C is a variant of NVLink optimized for chip-to-chip (rather than GPU-to-GPU) connectivity. NVLink-C2C supports cache coherence and different processor types (CPU, GPU, DPU), while standard NVLink connects only GPUs with non-coherent high-bandwidth links.
What systems use NVLink-C2C?
Grace Hopper GH200 superchip uses NVLink-C2C for CPU-GPU connectivity. Future NVIDIA platforms will extend NVLink-C2C to connect DPUs and additional processors in unified architectures.