16 January 2022

Multiresolution hash encoding cuts neural graphics training to seconds on a single GPU

Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller introduced a multiresolution hash encoding that trains neural graphics primitives in seconds and renders at 1920×1080 in tens of milliseconds, achieving a combined speedup of several orders of magnitude when the encoding and optimised CUDA kernels are used together.

Architectures and models Neural networks Deep learning Computer vision Foundational method Demonstrated

Background

Neural graphics primitives are neural networks trained to represent scenes, objects or other continuous signals, such as the colour at every point in a 3D volume. They produce high-quality results but come at a steep cost. As the authors put it, they ‘can be costly to train and evaluate’. That cost came from two places: the networks themselves are large, and every query into them demands many floating-point operations and memory accesses.

The standard fix was to give the network a positional encoding, converting coordinates into a richer set of features before feeding them in. But existing encodings did not scale well enough to close the gap between quality and speed.

What happened

Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller posted their paper to arXiv on 16 January 2022. Their solution was a ‘multiresolution hash table of trainable feature vectors whose values are optimized through stochastic gradient descent’. The idea is to store learned features at several spatial resolutions simultaneously, each level indexed by a hash of the input coordinates. A small neural network then reads from all of these tables and produces the final output.

Hash tables can have collisions, where two different positions map to the same entry. The multiresolution structure handles this gracefully: entries that collide at a coarse resolution are likely to be separated at a finer one, so ‘the network to disambiguate hash collisions, making for a simple architecture that is trivial to parallelize on modern GPUs’.

The team went further on the engineering side. Rather than relying on general-purpose GPU operations, they built ‘fully-fused CUDA kernels with a focus on minimizing wasted bandwidth and compute operations’.

The combined result was striking. The authors reported a combined speedup of several orders of magnitude, enabling training of high-quality neural graphics primitives in a matter of seconds, and rendering in tens of milliseconds at a resolution of 1920×1080.

The paper appeared at SIGGRAPH 2022, published in ACM Transactions on Graphics as volume 41, number 4, Article 102.

Why it mattered

Before this work, most neural graphics primitives required minutes to hours of training on high-end hardware. The method achieved a combined speedup of several orders of magnitude when the encoding and optimised CUDA kernels are used together, enabling training in seconds and rendering in tens of milliseconds at 1920×1080, making interactive neural rendering plausible on modern GPUs.

Sources

Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. arxiv.org. Primary source
Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. arxiv.org. Primary source

Cite this page

AI Achievements. (2022). Multiresolution hash encoding cuts neural graphics training to seconds on a single GPU. Retrieved 2026-08-29, from https://achievements.ai/milestone/multiresolution-hash-encoding-cuts-neural-graphics-training

@misc{achievements_multiresolution_hash_encoding_cuts_neural_graphics_training,
  title  = {Multiresolution hash encoding cuts neural graphics training to seconds on a single GPU},
  author = {{AI Achievements}},
  year   = {2022},
  url    = {https://achievements.ai/milestone/multiresolution-hash-encoding-cuts-neural-graphics-training}
}