NNI//SDK
Benchmarks Research Hub Docs ROI Calculator Sign InRegister Node
Benchmarks Research Hub Docs ROI Calculator Sign InRegister Node

Network Sockets were built for 1980s Ethernet.
NNI is built for 2025 Silicon.

Bypassing the Linux/Windows kernel for 0.0007ms intra-node latency. Zero-Copy. Zero-Tax. Neural-Native.

View the Rust Core on GitHub Read the Whitepaper (PDF)
pip install nni-sdk

The "Interconnect Tax" Comparison

Transparency in Performance. Tested on 2x NVIDIA H100 via PCIe Gen5 Bus.

Metric gRPC / Protobuf NVIDIA NCCL NNI-SDK
Data Path User-Kernel-User GPU-Direct Native L3/SRAM
Copy Cycles 3-4 Copies 1-2 Copies 0 Copies (Direct)
CPU Overhead High (12%+) Medium Near-Zero (<0.1%)
Latency (Avg) ~4.200 ms 0.050 ms 0.0007 ms

Live 4-Agent Cascade Simulation

Test the physical speed of direct memory routing versus standard sockets.

AWAITING

NNI Neural Cascade

0.0007 ms

Tail Latency (P99): 0.0009 ms

AWAITING

Standard TCP / JSON

1.1500 ms

Tail Latency (P99): 3.9100 ms

Estimated Cloud Infrastructure Savings: $4,250.00

Kernel-Bypass Integration

Deploy in 2 simple steps. No dependencies. Complete Hardware-Level execution.

Step 1: Install

pip install nni-sdk

Step 2: Run

nni_integration.py 
# 1. Initialize Global Hardware Slate
slate = nni_sdk.SharedSlate(name="global_numa_slate_01")

# 2. Extract active neural floats via Sparse Masking
packet = nni_sdk.Compressor.compress_sparse(
    floats=model_weights, 
    threshold=0.05
)

# 3. Stream directly to destination hardware via FFI -> Rust Core
# No Serialization. No Buffering. Pure Silicon Speed.
nni_sdk.Compressor.decompress_to_npu(packet, npu_sram_pointer)