US5091864A
Expired / Lapsed
US
⏱️ 0.0 years remaining
Systolic processor elements for a neural network
Assignee: Hitachi Ltd
•
Filing: 1989-12-21
•
Granted: 1992-02-25
•
Expires: 2009-12-21
Abstract
Abstract A neural net signal processor provided with a single layer neural net constituted of N neuron circuits which sums the results of the multiplication of each of N input signals Xj(j=1 to N) by a coefficient mij to produce a multiply-accumulate value ##EQU1## thereof, in which input signals Xj(j=1 to N) for input to the single layer neural net are input as serial input data, comprising: a multiplicity of systolic processor elements SPE-1(i=1 to M), each comprised of a two-state input data delay latch; a coefficient memory; means for multiplying and summing for multiply-accumulate output operations; an accumulator; a multiplexor for selecting a preceding stage multiply-accumulate output Sk(k=1 to i-1) and the multiply-accumulate product Si computed by the said circuit; wherein the multiplicity of systolic processor elements are serially connected to form an element array and element multiply-accumulate output operations are executed sequentially to obtain the serial multiply-accumulate outputs Si(i=1 to M) of one layer from the element array.
Inventors
- 👤 Toru Baji
- 👤 Hidenori Inouchi
Bibliographic Details
Application Number
—
Jurisdiction
US
Data Source
google_patents
Last Synchronized
2026-11-05 03:07
CPC Classifications
G; G06; G06N; G06N3/00; G06N3/02; G06N3/06; G06N3/063
Claims Analysis
Total Claims: 19
Claim 1
Independent Claim
270 words
A neural net signal processor provided with a single layer neural net including N neuron circuits, wherein N is any positive integer, which sum results of multiplication of each of N input signals X j (j=1 to N) by a coefficient m ij to produce a multiply-accumulate value ##EQU8## thereof, in which input signals X j (j=1 to N) for input to the single layer neural net are input as serial input data, comprising: a plurality of systolic processor elements SPE i (i=1 to M), wherein M is any positive integer, each systolic processor element including, storage means for storing coefficients, multiplying means for sequentially reading the coefficients from said storage means and sequentially multiplying each of the serial input data by an associated coefficient read out from said storage means, means for summing the results of the multiplications from said multiplying means to produce a multiply-accumulated value ##EQU9## a multiplexor for selectively outputting one of the multiply-accumulated values produced from said summing means and a multiply-accumulated value applied from a preceding systolic processor element, first latch for delaying the multiply-accumulated value applied from said multiplexor by a first time period corresponding to an operation period by at least one systolic processor element and for applying the delayed multiply-accumulated value to a succeeding systolic processor element, second latch for delaying the serial input data by a second time period longer than the first time period by an operation period by at least one systolic processor element and for applying the delayed multiply-accumulated value to the succeeding systolic processor element, whereby the single layer neural net produces a serial multiply-accumulated value.
Claim 2
Dependent Claim
270 words
A neural net signal processor provided with a single layer neural net including N neuron circuits, wherein N is any positive integer, which sum results of multiplication of each of N input signals X j (j=1 to N) by a coefficient m ij to produce a multiply-accumulate value ##EQU8## thereof, in which input signals X j (j=1 to N) for input to the single layer neural net are input as serial input data, comprising: a plurality of systolic processor elements SPE i (i=1 to M), wherein M is any positive integer, each systolic processor element including, storage means for storing coefficients, multiplying means for sequentially reading the coefficients from said storage means and sequentially multiplying each of the serial input data by an associated coefficient read out from said storage means, means for summing the results of the multiplications from said multiplying means to produce a multiply-accumulated value ##EQU9## a multiplexor for selectively outputting one of the multiply-accumulated values produced from said summing means and a multiply-accumulated value applied from a preceding systolic processor element, first latch for delaying the multiply-accumulated value applied from said multiplexor by a first time period corresponding to an operation period by at least one systolic processor element and for applying the delayed multiply-accumulated value to a succeeding systolic processor element, second latch for delaying the serial input data by a second time period longer than the first time period by an operation period by at least one systolic processor element and for applying the delayed multiply-accumulated value to the succeeding systolic processor element, whereby the single layer neural net produces a serial multiply-accumulated value.
Claim 3
Dependent Claim
67 words
A neural net signal processor according to claim 1, further provided with a second single layer neural net similarly constituted of M2 processor elements, wherein M2 is any positive integer, and in which the serial output of the said single layer neural net is used as input to the second single layer neural net and a multiply-accumulate output is obtained from the second single layer neural net.
Claim 4
Dependent Claim
102 words
A neural net signal processor according to claim 1, further provided with P-1 serially connected single layer neural nets, wherein P is any positive integer, similarly constituted of M2 processor elements, wherein M2 is any positive integer, and in which the multiply-accumulate output of the n-1the single layer neural net element array is used as input to the nth second single layer neural net and the serial output of the p-1th layer neural net element array is used as input to the element array of the pth layer neural net, whereby ultimately a multiply-accumulate output is obtained from the Pth neural net.
Claim 5
Dependent Claim
44 words
A neural net signal processor according to claim 3, wherein the element array of any one layer of P single layer neural nets is provided with an input data multiplexor that selects between re-inputting data into the element array and inputting using fresh data.
Claim 6
Dependent Claim
60 words
A neural net signal processor according to claim 3, wherein the element array of any one layer of P single layer neural nets is divided into two or more blocks and output from each block is output sequentially to a local data bus by a tri-state gate means and transferred to the element array of the next neural net layer.
Claim 7
Dependent Claim
35 words
A neural net signal processor according to claim 1, wherein the single layer neural net is provided with an input data multiplexor that selects between reinputting data into the element array and inputting fresh data.
Claim 8
Dependent Claim
38 words
A neural net signal processor according to claim 6, wherein outputs Y j (j=1 to M) of the element array of the single layer neural net is input to the element array and an output is obtained therefrom.
Claim 9
Dependent Claim
54 words
A neural net signal processor according to claim 1, wherein the single layer neural net element array is divided into two or more blocks and output from each block is output sequentially to a local data bus by a tri-state gate means and transferred to the element array of the next neural net layer.
Claim 10
Dependent Claim
111 words
A neural net signal processor according to claim 1, wherein each processor element is further provided with a nonlinear function look-up table and a multiplexor set* that bypasses the nonlinear function look-up table; a local data bus for bypassing data; a local input data bus for bypassing input data; gate means for outputting a multiply-accumulate output to the local data bus; gate means for obtaining input data from the local data bus or local input data bus; a multiplexor for selecting data from a preceding stage, bypassed data or data that has been fed back; and a multiplexor for selecting a multiply-accumulate output from a preceding stage or a prescribed input.
Claim 11
Dependent Claim
30 words
A neural net signal processor according to claim 1, in which the adder and multiplexor of each systolic processor elements are constituted as a parallel adder and a parallel multiplexor.
Claim 12
Dependent Claim
35 words
A neural net signal processor according to claim 1, in which the adder and multiplier of the systolic processor elements are constituted as a parallel adder and a series multiplier that uses the parallel adder.
Claim 13
Dependent Claim
54 words
A neural net signal processor according to claim 1 for processing P neurons, wherein P is any positive integer, and wherein each of the systolic processor elements also contains for P neuron coefficients m ij (i=1 to P, j=1 to N); an input hold register; P accumulators; P buffer memories; and P shift registers.
Claim 14
Dependent Claim
26 words
A neural net signal processor according to claim 1, wherein a shift register is employed for high speed read-out of coefficient mij from the coefficient memory.
Claim 15
Dependent Claim
29 words
A neural net signal processor according to claim 13, wherein coefficients in the coefficient memory are adapted to be rewritten by a host processor via a separate readout port.
Claim 16
Dependent Claim
33 words
A neural net signal processor according to claim 1, wherein the contents of the nonlinear function look-up table memory in each systolic processor element are adapted to be rewritten by the host processor.
Claim 17
Dependent Claim
69 words
A neural net signal processor according to claim 1 in which each systolic processor element is provided with an output data gate; bypass delay latch; test address bus; test data bus and an address decoder; and a test data generator/bypass control circuit; comparator; extended data memory and decision circuit are provided outside the systolic array; thereby enabling faults to be detected and a faulty processor element to be bypassed.
Claim 18
Dependent Claim
205 words
In a single layer neural net having a plurality of systolic processor elements connected in series for summing results of multiplication of each of serial input data by an associated coefficient to produce a multiply-accumulate value, each of said systolic processor elements comprising: storage means for storing coefficients; multiply-accumulating means for sequentially reading the coefficients from said storage means and sequentially multiplying each of the serial input data by an associated coefficient read out from said storage means to sum the results of the multiplications to produce a multiply-accumulated value; selection means for selectively outputting one of the multiply-accumulated value produced from said multiply-accumulating means and a multiply accumulated value applied from a preceding systolic processor element; first delay means for delaying the multiply-accumulated value applied from said selection means by a first time period corresponding to an operation period by at least one systolic processor element and for applying the delayed multiply-accumulated value to a succeeding systolic processor element; and second delay means for delaying the serial input data by a second time period larger than the first time period by an operation period by at least one systolic processor element and for applying the delayed multiply-accumulated value to a succeeding systolic processor element.
Claim 19
Dependent Claim
205 words
In a single layer neural net having a plurality of systolic processor elements connected in series for summing results of multiplication of each of serial input data by an associated coefficient to produce a multiply-accumulate value, each of said systolic processor elements comprising: storage means for storing coefficients; multiply-accumulating means for sequentially reading the coefficients from said storage means and sequentially multiplying each of the serial input data by an associated coefficient read out from said storage means to sum the results of the multiplications to produce a multiply-accumulated value; selection means for selectively outputting one of the multiply-accumulated value produced from said multiply-accumulating means and a multiply accumulated value applied from a preceding systolic processor element; first delay means for delaying the multiply-accumulated value applied from said selection means by a first time period corresponding to an operation period by at least one systolic processor element and for applying the delayed multiply-accumulated value to a succeeding systolic processor element; and second delay means for delaying the serial input data by a second time period larger than the first time period by an operation period by at least one systolic processor element and for applying the delayed multiply-accumulated value to a succeeding systolic processor element.
References Cited (Backward Citations)
US4953099A
Cited
US4967369A
Cited
USPTO Annuity & Maintenance Windows
| Annuity Stage | Window Opens | Payment Deadline | Surcharge Period Until | Estimated Fee | Status |
|---|---|---|---|---|---|
| 3.5 Years (Year 4) | 1995-02-24 | 1995-08-26 | 1996-02-25 | $2000.00 | Window Closed / Lapsed |
| 7.5 Years (Year 8) | 1999-02-24 | 1999-08-26 | 2000-02-25 | $3760.00 | Window Closed / Lapsed |
| 11.5 Years (Year 12) | 2003-02-24 | 2003-08-26 | 2004-02-25 | $7700.00 | Window Closed / Lapsed |