Neural Network Architecture

Overview

NIKOLA uses a state-of-the-art Efficiently Updatable Neural Network (NNUE) for position evaluation. Unlike traditional hand-crafted evaluation functions, NNUE learns complex positional patterns from millions of high-quality games and engine self-play.

HalfKA Architecture

NIKOLA employs the HalfKA (Half-King-All) feature set, which encodes:

• King position - Relative location of each side's king
• Piece placement - All pieces indexed relative to king position
• Perspective - Separate views for white and black

This representation enables efficient incremental updates as pieces move, requiring only sparse matrix operations rather than full network evaluation.

Network Topology

GPU-Batched Evaluation

NIKOLA's neural network inference is fully GPU-accelerated using native MIND tensor operations. The GPU-batched NNUE system collects positions from Lazy SMP search threads and evaluates them in batches for maximum throughput:

• Batched inference - 32-256 positions per GPU kernel, achieving 500M+ positions/sec
• Tensor cores - FP8/INT8 mixed precision on Blackwell and Vera Rubin architectures
• Multi-GPU distribution - Scale to 8+ GPUs per node with NVLink 5.0
• Async CUDA streams - Overlap compute with memory transfers for maximum utilization
• Virtual loss - Enable parallel MCTS expansion without tree corruption

Supported GPU Architectures

Incremental Updates

The NNUE architecture enables "efficiently updatable" evaluation. When a move is made, only the affected features need recalculation rather than the entire network. This provides 10-100x speedup compared to full evaluation, critical for deep search.

Training Data

NIKOLA's network is trained on a proprietary dataset of over 10 billion positions generated through:

• Self-play games at various time controls
• High-depth analysis of master games
• Curated endgame positions from tablebases
• Adversarial positions designed to expose weaknesses

setoption name NNUEPath value /path/to/custom.nnue