Our Patents

KoolBits is protected by patent-pending technology covering adaptive compression and direct computation of AI model parameter tensors.

Patent PendingU.S. Provisional Patent Application

Systems and Methods for Adaptive Compression and Processing of Model Parameter Tensors

Systems and methods for adaptive compression and direct computation of model parameter tensors in machine learning models. Tensor characteristics such as sparsity, statistical distribution, and access patterns are analyzed to select a compression strategy balancing storage, computation, and accuracy.

Inventors

Franz Torrez QuirogaLuc Pierre FaucheuxPeter Lanier Thomson

What Makes KoolBits Different

TREE/INDEX Recursive Compression

Recursively partitions parameter values into branches, each encoding a unique index. Produces collision-free compressed representations. Repeated branches are cached to reduce storage, and independent branches enable parallel execution on GPUs and multi-core processors.

Loss-Free Compression for LLMs

When patterns in operation sequences and index lists are identified in advance, the method achieves loss-free compression — no approximation error, no accuracy degradation. Especially suited for large language models with bounded values and long-range correlations.

Direct Computation on Compressed Data

Mathematical operations execute directly on compressed or indexed representations without full decompression. This reduces memory usage, accelerates computation, and improves energy efficiency across edge, data-center, and distributed platforms.

Implicit im2col Integration

Integrates the im2col operation directly within the matrix-multiplication kernel in the compressed domain. Eliminates intermediate matrices, reduces peak memory usage, and avoids preprocessing overhead on GPUs.

Why Existing Approaches Fall Short

Large language models comprise millions to billions of parameters stored as high-dimensional tensors. Conventional dense formats and fixed-precision storage scale poorly — producing excessive memory footprints, bandwidth bottlenecks, and elevated energy consumption.

Existing methods like pruning, quantization, and low-rank factorization either provide limited compression or disrupt critical statistical correlations. Most require full decompression before computation. Although small in isolation, approximation errors propagate across billions of tensor operations and produce deviations that reduce accuracy, coherence, or stability.

KoolBits addresses these limitations with a system that produces lossless, exact representations of neural network weights while enabling direct computational operations without decompressing each weight — applicable from edge devices to high-performance data centers.

Binary Multiplication as Organized Bitwise Processes

The disclosed method leverages the inherent properties of binary multiplication, wherein multiplication is realized through iterative summations combined with corresponding bit-shift operations — effectively reconstructing the multiplication function from fundamental binary manipulations. This enables operations on compressed data without the overhead of decompression.

50%+

Size reduction

0%

Accuracy loss

GPU

Parallel execution

Interested in Licensing KoolBits?

Contact our team to discuss enterprise licensing, partnerships, and OEM integration opportunities.

Frequently Asked Questions