18456235. COMPUTE OPTIMIZATIONS FOR LOW PRECISION MACHINE LEARNING OPERATIONS simplified abstract (Intel Corporation)

COMPUTE OPTIMIZATIONS FOR LOW PRECISION MACHINE LEARNING OPERATIONS

Organization Name

Intel Corporation

Inventor(s)

Elmoustapha Ould-ahmed-vall of Chandler AZ (US)

Sara S. Baghsorkhi of San Jose CA (US)

Anbang Yao of Beijing (CN)

Kevin Nealis of San Jose CA (US)

Xiaoming Chen of Shanghai (CN)

Altug Koker of El Dorado Hills CA (US)

Abhishek R. Appu of El Dorado Hills CA (US)

John C. Weast of Portland OR (US)

Mike B. Macpherson of Portland OR (US)

Dukhwan Kim of San Jose CA (US)

Linda L. Hurd of Cool CA (US)

Ben J. Ashbaugh of Folsom CA (US)

Barath Lakshmanan of Chandler AZ (US)

Liwei Ma of Beijing (CN)

Joydeep Ray of Folsom CA (US)

Ping T. Tang of Edison NJ (US)

Michael S. Strickland of Sunnyvale CA (US)

COMPUTE OPTIMIZATIONS FOR LOW PRECISION MACHINE LEARNING OPERATIONS - A simplified explanation of the abstract

This abstract first appeared for US patent application 18456235 titled 'COMPUTE OPTIMIZATIONS FOR LOW PRECISION MACHINE LEARNING OPERATIONS

Simplified Explanation

The abstract of the patent application describes a general-purpose graphics processing unit (GPU) that includes a dynamic precision floating-point unit. This unit has a control unit with precision tracking hardware logic to monitor the number of bits of precision for computed data in relation to a target precision. The dynamic precision floating-point unit also has computational logic to output data at multiple precisions.

• The patent application describes a general-purpose GPU with a dynamic precision floating-point unit.
• The dynamic precision floating-point unit includes a control unit with precision tracking hardware logic.
• The precision tracking hardware logic monitors the number of bits of precision for computed data.
• The precision tracking hardware logic compares the computed data precision to a target precision.
• The dynamic precision floating-point unit also includes computational logic to output data at multiple precisions.

Potential applications of this technology:

• Graphics processing: The dynamic precision floating-point unit can enhance the performance and efficiency of graphics processing in GPUs.
• Scientific computing: The ability to output data at multiple precisions can be beneficial for scientific simulations and calculations.
• Machine learning: GPUs are commonly used in machine learning applications, and the dynamic precision floating-point unit can improve the accuracy and efficiency of these computations.

Problems solved by this technology:

• Precision optimization: The precision tracking hardware logic helps optimize the precision of computed data, ensuring it meets the desired target precision without unnecessary over-precision.
• Performance improvement: By outputting data at multiple precisions, the dynamic precision floating-point unit can improve the overall performance of the GPU, especially in scenarios where different levels of precision are required.

Benefits of this technology:

• Improved efficiency: The dynamic precision floating-point unit allows for more efficient use of computational resources by dynamically adjusting the precision of computed data.
• Enhanced accuracy: The precision tracking hardware logic ensures that computed data meets the desired precision, resulting in more accurate calculations.
• Versatility: The ability to output data at multiple precisions makes the GPU suitable for a wide range of applications that require different levels of precision.

Original Abstract Submitted

One embodiment provides a general-purpose graphics processing unit comprising a dynamic precision floating-point unit including a control unit having precision tracking hardware logic to track an available number of bits of precision for computed data relative to a target precision, wherein the dynamic precision floating-point unit includes computational logic to output data at multiple precisions.