Topology View

A visual diagram of your GPU interconnect topology, showing how GPUs are connected via PCIe, NVLink, or the unified memory architecture on Apple Silicon.

Press g from any view to enter Topology mode.

gpulse Topology view showing GPU interconnect diagram with PCIe/NVLink connections and PRO pill badges
Topology view — GPU interconnect diagram with NVLink and scrollable content

What You See

Topology view renders a tree or graph diagram showing:

  • GPU Nodes — each GPU represented as a box with its index, model name, and current VRAM usage
  • Bus Connections — lines between GPUs showing the interconnect type (PCIe, NVLink, NVSwitch, Apple Unified Memory)
  • PCIe Root Complex — the CPU or IOH node that GPUs connect through
  • Link Bandwidth — theoretical bandwidth of each link (e.g., PCIe 4.0 x16 = 32 GB/s)
  • NUMA Node — which CPU socket each GPU is closest to (on multi-socket systems)

Interconnect Types

PCIe

Standard bus for connecting GPUs. Common configurations: PCIe 3.0 x16 (16 GB/s), PCIe 4.0 x16 (32 GB/s), PCIe 5.0 x16 (64 GB/s). P2P transfers between GPUs on the same root complex are faster than those crossing CPU sockets.

NVLink

NVIDIA's high-bandwidth GPU-to-GPU link. NVLink 3.0 provides 600 GB/s bidirectional per link. Used in DGX systems, A100, H100, and other data center GPUs. Enables much faster multi-GPU communication than PCIe.

Apple Unified Memory

On Apple Silicon, the GPU and CPU share the same memory pool. There is no discrete GPU bus — the "interconnect" is the unified memory architecture itself, with bandwidth determined by the SoC (e.g., M4 Max = 546 GB/s).

Why Topology Matters

Understanding GPU topology is critical for:

  • Multi-GPU Training — frameworks like PyTorch and DeepSpeed can optimize communication patterns based on topology. GPUs connected via NVLink should be grouped together.
  • NUMA Awareness — for best performance, allocate CPU memory on the same NUMA node as the GPU. Topology view shows which socket each GPU belongs to.
  • Debugging Slow All-Reduce — if gradient synchronization is slow, check whether your GPUs are on separate root complexes or crossing QPI links.

Keyboard Shortcuts

Key Action
Tab / Shift+Tab Select next / previous GPU node
d Open selected GPU in Detail view
v Return to Grid view

When to Use Topology View

Use Topology view when setting up a new multi-GPU machine, debugging communication bottlenecks in distributed training, or verifying hardware configuration after changes. It is read-only and does not require any active workload.