Nvidia describes an approach to numerical operations between neural-network layers: Shared scales are intended to support element-wise addition and subtraction at reduced precision. US application US20260300712A1 was published on October 1, 2026. It presents a calculation method, not a new AI model.
Why using the same scale matters
Quantization maps numbers onto a coarser grid. When two inputs use different scales, their compact values cannot simply be added. Nvidia therefore determines shared scaling factors, applies them to the inputs and then performs the element-wise operation. The factors can cover entire inputs or corresponding tensor blocks. They are calculated in advance from representative data or dynamically. Between candidates derived from input and output statistics, the method can select the scale with the lower mean squared error.
The original document describes the process in independent claim 1, calibration in claims 7–13 and shared factors for subregions in claims 17–18.
What evidence is missing
The document provides no model benchmarks, measured speedups or energy savings. It does not establish a particular bit width or product integration. Whether shared scales reduce resource requirements in practice also depends on rounding errors and calibration overhead. Comparisons against higher precision on the same models and hardware would be needed.
Document and status
“Scaling Element-Wise Operations of a Neural Network”: Applicant NVIDIA Corporation; US application 19/091,697, filed March 26, 2025. The original document lists no earlier priority. A1 denotes an application publication, not a patent grant. We could not independently confirm the current official prosecution status on October 3, 2026.
Pandorex View: The interesting point is the computation between network components. If that can also be quantized reliably, more processing steps could use compact numbers. The actual benefit remains a question for measurement.