nCPU is a revolutionary CPU architecture that operates entirely on GPU, leveraging the power of tensors for registers, memory, flags, and the program counter. Each arithmetic operation is executed through a trained neural network, allowing for innovative computational techniques that optimize performance and efficiency. Key features of nCPU include:

Core Features

• Neural Network-Based ALU: All arithmetic operations are executed via neural networks, enabling adaptive calculations without hardcoded arithmetic. For example, addition employs Kogge-Stone carry-lookahead and multiplication utilizes a learned byte-pair lookup table.

• Performance Optimization: The architecture utilizes attention-based routing for shifts and neural truth tables for bitwise operations, achieving remarkable accuracy in integer computations and verified through extensive automated tests.

• Comprehensive Model Inventory: nCPU comes with 23 trained models, with 13 actively wired into operations, providing 100% accuracy across a variety of computational tasks. Each model is methodically designed for specific instructions, including arithmetic and logical operations.

How nCPU Works

The architecture of nCPU is built to maximize performance, with all critical components residing on the GPU. This includes:

• Registers, Memory, and Flags: All states are maintained in GPU-resident tensors, ensuring rapid access and reduced latency.
• Execution Modes: The CPU operates in two distinct modes:
  • Neural Mode: Executes each operation through a forward pass in a trained model.
  • Fast Mode: Utilizes native tensor operations for maximum throughput.

Performance Metrics

nCPU has displayed impressive benchmarks, particularly in operations like multiplication, which distinctly outpaces addition thanks to its neural architecture.

• Latency: While traditional CPUs often have slower multiplication operations, nCPU achieves speeds of ~21 microseconds for multiplications due to its unique design.
• Efficiency: The architecture manages to handle multiple operations efficiently, showcasing the potential of neural networks in executing CPU functions traditionally held by hardware.

Research and Development

Extensive testing and benchmarking have validated the architecture's effectiveness, demonstrating a significant reduction in execution time for various operations over traditional methods. A detailed analysis can be found in the accompanying research paper.

Use Cases

Recent demos, such as the DOOM raycaster, highlight nCPU's capability, running all arithmetic through neural models while maintaining coherent output, whether in neural or fast mode. Users can explore further through various assembly programs that illustrate the potential of this pioneering CPU innovation.

For developers and researchers interested in exploring cutting-edge computational architectures, nCPU offers a unique platform that integrates state-of-the-art machine learning models into traditional CPU functionality.