neuratensor 1.1.3

Production neuromorphic inference SDK with CUDA acceleration

NeuraTensor SDK

Production neuromorphic inference at 8ms latency

NeuraTensor is a high-performance neuromorphic inference runtime featuring a proprietary fused SNN-SSM architecture with CUDA acceleration. Optimized for NVIDIA Jetson AGX Orin and edge deployment.

---

⚡ Performance

• 8.13ms mean latency (Jetson AGX Orin, batch=1, seq=64, FP16)
• 123 sequences/sec throughput
• Sub-10ms real-time inference guarantee
• Up to 120× lower latency vs standard PyTorch execution (measured on Jetson AGX Orin, batch=1, seq=64, same model architecture)

---

🚀 Quick Start

Installation

pip install neuratensor

Python API

from neuratensor import NeuraTensor, NeuraTensorConfig
import torch

# Load model (3 lines)
config = NeuraTensorConfig.preset("64m")
model = NeuraTensor(config).cuda().half()

# Inference (1 line)
input_ids = torch.randint(0, 50000, (1, 64), device="cuda")
output = model(input_ids)
# ✅ 8ms latency, 64M params, 123 seq/s

CLI Benchmark

# Run benchmark
neuratensor benchmark 64m --iterations 100

# Example output:
# ============================================================
# NeuraTensor 64M | Benchmark
# ============================================================
# Device:    Orin (SM 8.7)
# CUDA:      11.4
# ------------------------------------------------------------
# Latency (mean):    8.13 ms ± 0.07 ms
# Latency (p99):     8.22 ms
# Throughput:        123.0 seq/s
# Kernel:            fused_snn_ssm (secure)
# ============================================================

---

🏗️ Architecture

• Hybrid SNN-SSM: 6 spiking layers + 10 state space layers
• Fused CUDA kernels: Single-pass execution (patent pending)
• FP16 optimized: Hardware tensor cores
• Deterministic latency: No dynamic dispatch overhead

---

📦 Available Models

Model	Parameters	Latency (Orin)	Throughput	Use Case
64M	64,155,526	8.13ms	123 seq/s	Edge, real-time
256M	256M	~15ms	67 seq/s	Balanced
1B	1B	~35ms	29 seq/s	Quality

All measurements: Jetson AGX Orin, batch=1, seq=64, FP16

---

⚠️ What This Is NOT

NeuraTensor is NOT:

• ❌ A training framework
• ❌ A general-purpose Transformer library
• ❌ A drop-in replacement for PyTorch
• ❌ An autograd-enabled tensor library

NeuraTensor IS:

• ✅ A production-grade inference runtime
• ✅ Optimized for edge latency and determinism
• ✅ Purpose-built for neuromorphic architectures
• ✅ A compiled binary SDK with Python bindings

---

💻 Supported Hardware

Tested & Validated:

• ✅ Jetson AGX Orin (primary target)
• ✅ Jetson Orin NX
• ✅ Jetson Orin Nano (limited testing)

Experimental Support:

• ⚠️ NVIDIA RTX 30xx / 40xx (desktop GPUs)
• ⚠️ A100, H100 (datacenter GPUs)

Requirements:

• CUDA Compute Capability: 7.0 or higher (SM 7.0+)
• CUDA Toolkit: 11.4 or higher
• Driver: 470.x or higher
• Python: 3.8, 3.9, 3.10
• PyTorch: 2.0+ (for tensor compatibility only)

Note: If your GPU is unsupported, the SDK will raise a clear error at import time. Source code for the core runtime is not distributed.

---

📦 Binary Distribution Model

The SDK ships with precompiled CUDA binaries for maximum performance:

• ✅ Optimized for Jetson AGX Orin (ARM64 + CUDA 11.4)
• ✅ No compilation required at install time
• ✅ Symbols obfuscated for IP protection
• ⚠️ Platform-specific (Linux ARM64 only for v1.x)

If your system configuration is unsupported, you will see:

RuntimeError: CUDA kernel not compatible with this device

Source code for the proprietary core is not distributed. Binary-only distribution is intentional.

---

🔒 Security & IP Notice

NeuraTensor contains proprietary CUDA kernels and runtime logic protected by:

• Patent-pending architecture (USPTO filing: pending)
• Obfuscated binaries (symbols stripped and renamed)
• Restricted license (see PATENT_NOTICE.txt)

Reverse engineering, redistribution, or modification of the core binaries is prohibited by the license.

For commercial licensing, custom hardware support, or source code access:

• Email: [e-mail]
• Website: https://neuramorphic.ai

---

🔧 Requirements

System:

• OS: Linux (Ubuntu 20.04+ recommended)
• Architecture: ARM64 (Jetson) or x86_64 (experimental)
• CUDA: 11.4+ with cuDNN
• GPU Memory: 2GB minimum

Python:

pip install neuratensor

# Dependencies (auto-installed):
# - torch >= 2.0.0
# - numpy >= 1.19.0

---

📚 Documentation

• Quick Start: examples/
• API Reference: Coming soon
• Performance Guide: Coming soon
• Hardware Guide: Coming soon

---

🤝 Support

Community:

• GitHub Issues: For bugs and feature requests (repo link TBD)
• PyPI Page: https://pypi.org/project/neuratensor/

Enterprise:

• Custom Models: [e-mail]
• Licensing: [e-mail]
• NVIDIA Partners: [e-mail]

---

📊 Citation

If you use NeuraTensor in your research or product, please cite:

@software{neuratensor2025,
  title = {NeuraTensor: High-Performance Neuromorphic Inference SDK},
  author = {Neuramorphic, Inc.},
  year = {2025},
  url = {https://pypi.org/project/neuratensor/}
}

---

📄 License

Proprietary License - Binary distribution only.

• ✅ Evaluation and non-commercial use permitted
• ⚠️ Commercial use requires separate license agreement
• ❌ Reverse engineering strictly prohibited
• ❌ Redistribution not permitted

See LICENSE and PATENT_NOTICE.txt for details.

---

🏢 About

NeuraTensor is developed by Neuramorphic, Inc.

Patent-pending neuromorphic architecture combining spiking neural networks with state space models for ultra-low latency inference on edge devices.

---

© 2024-2025 Neuramorphic, Inc. All rights reserved.