9
4.2
System Architecture
NF5448A6 uses a balanced CPU-GPU connection topology.
Figure 4-2 NF5448A6 System Architecture
The uplink bandwidth of CPU-to-GPU connections is provided via 4 PCIe 4.0 x16 cards.
This greatly increases the bandwidth for data communication between CPUs and GPUs
to prevent bottlenecks in AI training involving a large amount of data. NVIDIA
®
NVLink
3.0 is supported between GPUs, with up to 600 GB/s of P2P interconnection bandwidth.
The NVLink
™
full-interconnection topology reduces latency in case of data sharing
among multiple GPUs, thereby providing more efficient computing performance.