Intel Corporation

Bogdan Mihai Pasca

Martin Langhammer

Gregg William Baeckler

Sergey Vladimirovich Gribok

Dmitry Nikolai Denisenko

Andrei Hagiescu

An integrated circuit device includes multiplier circuitry configured to determine a plurality of columns of subproducts by multiplying a plurality of values. Each column of the plurality of columns includes one or more subproducts of a plurality of subproducts. The integrated circuit device also includes adder circuitry configured to determine a plurality of sums, each sum being a sum of one column of the plurality of columns. A first portion of the adder circuitry associated with a first column of the plurality of columns is configured to receive a first value and second value that are associated with the first column and a third value associated with a second column of the plurality of columns that differs from the first column. The third value is a carry-out value generated by a second portion of the adder circuitry associated with the second column of the plurality of columns.

Systems and methods for low latency modular multiplication

The present disclosure relates generally to techniques for adjusting the number representation (e.g., format) of a variable before and/or after performing one or more arithmetic operations on the variable. In particular, the present disclosure relates to scaling the range of a variable to a suitable representation based on available hardware (e.g., hard logic) in an integrated circuit device. For example, an input in a first number format (e.g., bfloat16) may be scaled to a second number format (e.g., half-precision floating-point) so that circuitry implemented to receive inputs in the second number format may perform one or more arithmetic operations on the input. Further, the output produced by the circuitry may be scaled back to the first number format. Accordingly, arithmetic operations, such as a dot-product, performed in a first format may be emulated by scaling the inputs to and/or the outputs from arithmetic operations performed in another format.

Floating-point dynamic range expansion

Folded integer multiplier (FIM) circuitry includes a multiplier configurable to perform multiplication and a first addition/subtraction unit and a second addition/subtraction unit both configurable to perform addition and subtraction. The FIM circuitry is configurable to determine each product of a plurality of products for a plurality of pairs of input values having a first number of bits by performing, using the first and second addition/subtraction units, a plurality of operations involving addition or subtraction, and performing, using the multiplier, a plurality of multiplication operations involving values having fewer bits than the first number of bits. The plurality of multiplication operations includes a first number of multiplication operations, and the multiplier is configurable to begin performing all multiplication operations of the plurality of multiplication operations within a first number of clock cycles equal to the first number of multiplication operations.

Folded integer multiplication for field-programmable gate arrays

The present disclosure relates generally to techniques for enhancing adders implemented on an integrated circuit. In particular, arithmetic performed by an adder implemented to receive operands having a first precision is restructured so that a set of sub-adders performs the arithmetic on a respective segment of the operands. More specifically, the adder is restructured, and a decoder determines a generate signal and a propagate signal for each of the sub-adders and routes the generate signal and the propagate signal to a prefix network. The prefix network determines respective carry bit(s), which carries into and/or select a sum at a subsequent sub-adder.

Prefix network-directed addition

An integrated circuit with a large multiplier is provided. The multiplier may be configured to receive large input operands with thousands of bits. The multiplier may be implemented using a multiplier decomposition scheme that is recursively flattened into multiple decomposition levels to expose a tree of adders. The adders may be collapsed into a merged pipelined structure, where partial sums are forwarded from one level to the next while bypassing intervening prefix networks. The final correct sum is not calculated until later. In accordance with the decomposition technique, the partial sums are successively halved, which allows the prefix networks to be smaller from one level to the next. This allows all sums to be calculated at approximately the same pipeline depth, which significantly reduces latency with no or limited pipeline balancing.

Reduced latency multiplier circuitry for very large numbers

The present disclosure relates generally to techniques for enhancing recurrent neural networks (RNNs) implemented on an integrated circuit. In particular, approximations of activation functions used in an RNN, such as sigmoid and hyperbolic tangent, may be implemented in an integrated circuit, which may result in increased efficiencies, reduced latency, increased accuracy, and reduced resource consumption involved with implementing machine learning.

Hyperbolic functions for machine learning acceleration

An integrated circuit is provided with a modular multiplication circuit. The modular multiplication circuit includes an input multiplier for computing the product of two input signals, truncated multipliers for computing another product based on a modulus value and the product, and a subtraction circuit for computing a difference between the two products. An error correction circuit uses the difference to look up an estimated quotient value and to subtract out an integer multiple of the modulus value from the difference in a single step, wherein the integer multiple is equal to the estimated quotient value. A final adjustment stage is used to remove any remaining residual estimation error.

Integrated circuits with modular multiplication circuitry

A programmable device may be configured to support machine learning training operations using matrix multiplication circuitry implemented on a systolic array. The systolic array includes an array of processing elements, each of which includes hybrid floating-point dot-product circuitry. The hybrid dot-product circuitry has a hard data path that uses digital signal processing (DSP) blocks operating in floating-point mode and a hard/soft data path that uses DSP blocks operating in fixed-point mode operated in conjunction with general purpose soft logic. The hard/soft data path includes 2-element dot-product circuits that feed an adder tree. Results from the hard data path are combined with the adder tree using format conversion and normalization circuitry. Inputs to the hybrid dot-product circuitry may be in the BFLOAT16 format. The hard data path may be in the single precision format. The hard/soft data path uses a custom format that is similar to but different than BFLOAT16.

Machine learning training architecture for programmable devices

Adder trees may be constructed for efficient packing of arithmetic operators into an integrated circuit. The operands of the trees may be truncated to pack an integer number of nodes per logic array block. As a result, arithmetic operations may pack more efficiently onto the integrated circuit while providing increased precision and performance.

Reduction operation mapping systems and methods

Integrated circuits with digital signal processing (DSP) blocks are provided. A DSP block may include one or more large multiplier circuits. A large multiplier circuit (e.g., an 18&#xd7;18 or 18&#xd7;19 multiplier circuit) may be used to support two or more smaller multiplication operations sharing one or two sets of multiplier operands, a complex multiplication, and a sum of two multiplications. If the multiplier products overflow and interfere with one another, correction operations can be performed. Partial products from two or more larger multiplier circuits can be used to combine decomposed partial products. A large multiplier circuit can also be used to support two floating-point mantissa multipliers.

Methods for using a multiplier to support multiple sub-multiplication operations

Integrated circuits with digital signal processing (DSP) blocks are provided. A DSP block may include one or more large multiplier circuits. A large multiplier circuit such as an 18&#xd7;18 multiplier circuit may be used to support two or more smaller multiplication operations such as two 8&#xd7;8 integer multiplications or two 9&#xd7;9 integer multiplications. To implement the two 8&#xd7;8 or 9&#xd7;9 unsigned/signed multiplications, the 18&#xd7;18 multiplier may be configured to support two 8&#xd7;8 multiplications with one shared operand, two 6&#xd7;6 multiplications without any shared operand, or two 7&#xd7;7 multiplications without any shared operand. Any potential overlap of partial product terms may be subtracted out using correction logic. The multiplication of the remaining most significant bits can be computed using associated multiplier extension logic and appended to the other least significant bits using merging logic.

Growing community of inventors

Toulouse, France

Bogdan Mihai Pasca