Skip to main content
GPU Glossary/Networking
Networking

SHARP (Scalable Hierarchical Aggregation and Reduction Protocol)

In-network compute for accelerating collective operations.

What it is

SHARP (Scalable Hierarchical Aggregation and Reduction Protocol) is an NVIDIA Networking technology that offloads collective operations like AllReduce directly to InfiniBand switches, performing data aggregation in the network fabric rather than at endpoints. It requires compatible InfiniBand switch firmware and careful tree configuration.

Why it matters

SHARP can reduce AllReduce latency by up to 2x and free GPU compute cycles from reduction math, directly improving training throughput for data-parallel workloads where AllReduce is on the critical path. When SHARP is misconfigured or a switch firmware upgrade breaks compatibility, NCCL silently falls back to endpoint-based reductions with no error -- throughput degradation is the only signal.

How to monitor

Verify SHARP is active via NCCL_DEBUG=INFO logs and check for SHARP tree resource allocation in the subnet manager logs. Monitor IB switch firmware versions for SHARP compatibility after any upgrade. Factryze tracks NCCL collective throughput as a proxy for SHARP health and flags unexpected inter-node AllReduce latency increases.

SHARP - In-Network AggregationSHARP - In-Network Aggregation
Pinch to zoom, drag to pan, double-tap to toggle
SHARP - In-Network AggregationSHARP - In-Network Aggregation

Related terms

InfiniBand (Networking)

High-bandwidth, low-latency network fabric for GPU clusters.

Adaptive Routing (Networking)

Dynamic path selection in network switches to avoid congestion.

Network Fabric (Networking)

The physical interconnect topology connecting all nodes in a cluster.

Monitor this automatically

Factryze correlates GPU signals in real time: errors, clocks, and fabric health.

Get Started Free