Convolution memory
AssigneeSteven G. Morton
InventorSteven G. Morton
Life of this patent
Estimating the term…
Loading patent…
AssigneeSteven G. Morton
InventorSteven G. Morton
Estimating the term…
Method and apparatus including a system of integrated ciruit devices for executing matrix operations of the form [A] operated upon by [B] equals [C], where [A], [B] and [C] are each a matrix having a plurality of elements expressed in either unsigned or in two's complement format. The system includes at least one first array of data storage means organized as (j) physical rows and (k) physical columns of bits for providing storage for one bit of each of a plurality of elements of a row vectors of the matrix [A], an individual one of the (j) rows having bits of the same binary weight. The system further includes at least one second array of data storage means organized as (m) physical rows and (n) physical columns of bits, an individual one of the (m) rows storing a same binary weight bit or each of a plurality of elements of a column vector of the matrix [B]. The system further includes a processing device for processing in a predetermined manner a specified one of the (j) rows of the first array of data storage means with a specified one of the (m) rows of the second array of data storage means for producing a plurality of (x) output bits. The system also includes logic for determining a number of the output bits having a specified logic state and logic for determining a dot product.
A convolution memory chip comprising:
An integrated circuit device for executing matrix operations of the form {A} operated upon by {B} equals {C}, where {A}, {B} and {C} are each a matrix having a plurality of elements expressed in either unsigned or in two's complement format, comprising:
Apparatus for multiplying a matrix, said matrix having R rows and C columns of matrix elements each represented by MP matrix precision bits in either unsigned or in two's complement notation, by a column-vector having C rows of column-vector elements each represented by VP vector precision bits in either unsigned or in two's complement notation, said apparatus comprising a plurality of integrated circuit devices for storing and for processing at least the matrix and the column-vector and for outputting a result of a multiplication therebetween, said apparatus further comprising means, coupled to said devices, for controlling the operation thereof to sequentially process rows of the matrix, an individual one (N) of said integrated circuit devices comprising:
A method of multiplying a matrix times a vector, comprising the steps of:
A row-serial, bit serial, element-parallel method for multiplying a matrix times a vector with minimum latency, comprising the steps of:
A bit-serial, row-serial, element-parallel method for multiplying a matrix times a vector, comprising the steps of:
Dependencies are inferred from the claim text.