Multiple Double Arithmetic on NVIDIA Tensor Cores

arXiv:2607.06881 2026 Training 1 ideas extracted · analyzed Aug 30, 2026

What the math gives to ML

The paper presents a four-limb multiple-double representation whose multiplication is implemented as a convolution of limb products. Its transferable asset is the separation of fast tensor-core matrix multiplications from branch-heavy renormalization, allowing extra numerical precision to be retained across long reductions. A practical neural-network use is a selective high-precision path for numerically fragile matrix products or optimizer accumulators rather than replacing all low-precision computation.

Ideas from this paper

Unverified 2026

Tensor-Core Limb Expansion for Stable Accumulation

Represent selected activations, weights, or optimizer accumulators as four floating-point limbs and evaluate products through tensor-core matrix multiplications encoding limb convolution. Retain the convolution components during reductions and renormalize only at block boundaries, avoiding branch-heavy multi-double arithmetic inside every multiply-add.

Useful5/10
Difficulty6/10
Novelty6/10
Paper: Multiple Double Arithmetic on NVIDIA Tensor Cores arXiv:2607.06881