US 11,669,933 B2
Dynamic precision management for integer deep learning primitives
Naveen K. Mellempudi, Bangalore (IN); Dheevatsa Mudigere, Bangalore (IN); Dipankar Das, Pune (IN); and Srinivas Sridharan, Bangalore (IN)
Assigned to Intel Corporation, Santa Clara, CA (US)
Filed by Intel Corporation, Santa Clara, CA (US)
Filed on Apr. 27, 2022, as Appl. No. 17/730,364.
Application 16/853,405 is a division of application No. 15/881,991, filed on Jan. 29, 2018, granted, now 10,643,297, issued on May 5, 2020.
Application 17/730,364 is a continuation of application No. 17/083,588, filed on Oct. 29, 2020, granted, now 11,321,805.
Application 17/083,588 is a continuation of application No. 16/853,405, filed on Apr. 20, 2020, granted, now 10,825,127, issued on Nov. 3, 2020.
Claims priority of provisional application 62/501,796, filed on May 5, 2017.
Prior Publication US 2022/0327656 A1, Oct. 13, 2022
This patent is subject to a terminal disclaimer.
Int. Cl. G06T 1/20 (2006.01); G06F 5/01 (2006.01); G06F 7/501 (2006.01); G06F 7/523 (2006.01); G06F 7/544 (2006.01); G06F 17/15 (2006.01); G06F 17/16 (2006.01); G06N 3/063 (2023.01); G06N 3/084 (2023.01); G06N 3/044 (2023.01); G06N 3/045 (2023.01)
CPC G06T 1/20 (2013.01) [G06F 5/01 (2013.01); G06F 7/501 (2013.01); G06F 7/523 (2013.01); G06F 7/5443 (2013.01); G06F 17/153 (2013.01); G06F 17/16 (2013.01); G06N 3/044 (2023.01); G06N 3/045 (2023.01); G06N 3/063 (2013.01); G06N 3/084 (2013.01); G06F 2207/382 (2013.01); G06F 2207/4824 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A graphics processing unit comprising:
an interface to an interconnect fabric; and
a compute unit coupled with the interface, the compute unit including circuitry configured to:
quantize elements of a floating-point tensor to convert the floating-point tensor into a dynamic fixed-point tensor, wherein to quantize an element of the floating-point tensor, the circuitry is to compute a right-shift value based on a difference between an exponent value of the element of the floating-point tensor and the exponent value of an absolute maximum value of the floating-point tensor and right-shift a mantissa of the element based on the right-shift value to generate a magnitude integer;
perform a compute operation on input including the dynamic fixed-point tensor; and
generate an output tensor via the compute operation.