DecaGEMM · Efficient AI computeStructure-aware compute
Identify reusable mathematical structure and eliminate redundant work.
Efficient AI infrastructure
More compute. Less power.
DecaGEMM accelerates the matrix-multiplication workload at the core of today’s AI — increasing useful work per joule while preserving exact output correctness.
117.18×more logical work per joule*
~54 Wsame GPU power envelope*
99.15%less energy per logical step*
Software-onlyGEMM acceleration path
Training + inferenceone deterministic engine
Bit-identicalverified mathematical output*
Drop-in directiondesigned for current AI stacks
The energy compute wallAI does not only need more compute.
AI does not only need more compute.
It needs more Compute Capacity per Watt
Conventional acceleration optimizes hardware, kernels, precision or scheduling — while the underlying matrix-multiplication work remains.
DecaGEMM changes how that work is executed. DecaGEMM uses structure-aware, energy-based convergence to reduce data movement and arithmetic while preserving the required result.
DecaGEMM · The product
A new matrix-multiplication engine for current AI workloads.
Designed as a software-only acceleration path for enterprise, cloud and AI infrastructure.
AI workloadInput matrices
DecaGEMMEnergy convergence
AI workloadExact output
Less data movement
Reduce the matrix representation and the work that must move through memory.
Elastic depth
Compute only what is required and stop when the result is proven.
Deterministic correctness
Preserve exact output behavior without conventional FP accumulation inside GEMM.
Measured performance
More logical work from nearly the same instantaneous power.
RTX 5070 Laptop GPU · 32,768 logical tasks · 192 unique groups · burst depth 128 · 16 schedules per plan.
117.18×maximum measured logical work per joule
The breakthrough is not lower instantaneous GPU power. It is dramatically more useful computation from each joule.
median processing speed
29,533 jobs/s vs. 3,108 jobs/s across 40 paired tests.
lower ENERGY vs. cuBLAS
7.5895× NVIDIA/DecaGEMM energy ratio in the measured setup.
mismatches
Across 104,857,600 verified outputs in the measured campaign.
fewer calculations
6,291,264 vs. 536,854,528 scalar operations with bit-identical reconstruction.
* Internal measured benchmarks from the supplied August 2026 technical deck. Results are configuration- and workload-dependent and require external validation before generalized production claims.
Designed for adoption
Meet the current stack where it already runs.
DecaGEMM is being productized for evaluation without changing the model or adding specialized hardware.
Designed as a cuBLAS-compatible DLL path for familiar AI deployment workflows.
One deterministic GEMM mechanism across both major AI compute modes.
Evaluate throughput, energy per completed job and exact output under a fixed protocol.
Save energy now
Request an evaluation