VCSEL-based coherent scalable deep learning
The exponential growth in deep learning models is challenging existing computing hardware. Optical neural networks (ONNs) accelerate machine learning tasks with potentially ultrahigh bandwidth and nearly no loss in data movement. Scaling up ONNs involves improving scalability, energy efficiency, compute density, and inline nonlinearity. However, realizing all these criteria remains an unsolved challenge. Here, we demonstrate a three-dimensional spatial time-multiplexed ONN architecture based on dense arrays of microscale vertical cavity surface emitting lasers (VCSELs). The VCSELs, coherently injection-locked to a leader laser, operate at gigahertz data rates with a 7T-phase-shift voltage on the 10-millivolt level. Optical nonlinearity is incorporated into the ONN with no added energy cost using coherent detection of optical interference between VCSELs.
1 . An optical neural network comprising:
an array of vertical-cavity surface-emitting lasers (VCSELs) comprising a first VCSEL to emit a first beam phase-modulated with an activation vector and second VCSELs to emit second beams phase-modulated with weights of a weight matrix of the optical neural network and coherent with the first beam;
a diffractive optical element, in optical communication with the first VCSEL, to fan out the first beam; and
an array of photodetectors, in optical communication with the array of VCSELs and the diffractive optical element, to detect interference between respective fanned-out copies of the first beam from the diffractive optical element and the second beams.
2 . The optical neural network of claim 1 , wherein the array of VCSELs is monolithically integrated with the diffractive optical element and the array of photodetectors.
3 . The optical neural network of claim 1 , wherein the array of VCSELs is configured to be modulated with a half-wave voltage of less than 10 mV.
4 . The optical neural network of claim 1 , wherein the array of VCSELs is configured to be modulated at a rate of at least 1 Gb/s.
5 . The optical neural network of claim 1 , wherein the array of photodetectors is configured to generate outputs proportional to
∑
t
=
1
k
A
W
,
k
sin
(
ϕ
W
,
k
-
ϕ
X
)
,
where A X and ϕ X are the amplitude and phase, respectively, of the first beam and A W,k and ϕ W,k are the amplitude and phase, respectively, of the kth second beam.
6 . The optical neural network of claim 1 , wherein the optical neural network has a compute density of at least 10 TOPS/(mm 2 ·s).
7 . The optical neural network of claim 1 , wherein the optical neural network is configured to operate at an energy consumption of 1 fJ/OPS.
8 . The optical neural network of claim 1 , further comprising:
a leader laser, in optical communication with the array of VCSELs, to injection lock the array of VCSELs.
9 . The optical neural network of claim 1 , wherein the diffractive optical element is a first diffractive optical element, and further comprising:
a second diffractive optical element, in optical communication with one of the second VCSELs in the array of VCSELs, to fan out a second beam emitted by the one of the second VCSELs.
10 . The optical neural network of claim 1 , wherein the array of VCSELs, the diffractive optical element, and the array of photodetectors form one of a sequence of photonic tensor cores in the optical neural network.
11 . The optical neural network of claim 10 , further comprising:
an analog-to-digital converter (ADC), operably coupled to the array of photodetectors, to convert analog outputs of the array of photodetectors into serialized digital values;
a memory, operably coupled to the ADC, to store the serialized digital values; and
a digital-to-analog converter (DAC), operably coupled to the memory, to convert the serialized digital values into an analog input suitable for driving a VCSEL in a subsequent photonic tensor core in the sequence of photonic tensor cores.
12 . The optical neural network of claim 1 , wherein the first VCSEL is configured to phase-modulate the first beam with the activation vector using phase shift keying.
13 . The optical neural network of claim 1 , wherein the array of VCSELs is configured to phase-modulate the first beam and the second beams using thermo-optic modulation.
14 . The optical neural network of claim 1 , wherein the array of VCSELs is configured to phase-modulate the first beam and the second beams using free-carrier injection.
15 . The optical neural network of claim 1 , wherein the array of photodetectors is configured as an array of balanced homodyne receivers.
16 . The optical neural network of claim 1 , further comprising:
driver circuitry electrically and physically coupled to the array of VCSELs;
a first layer of transparent polymer bonding the array of VCSELs to the diffractive optical element; and
a second layer of transparent polymer bonding the diffractive optical element to the array of photodetectors.
17 . An optical neural network comprising:
an array of vertical-cavity surface-emitting lasers (VCSELs) comprising a first VCSEL to emit a first beam phase-modulated with an activation vector and second VCSELs to emit second beams phase-modulated with weights of a weight matrix of the optical neural network and coherent with the first beam;
a diffractive optical element, in optical communication with the first VCSEL, to fan out the first beam; and
an array of photodetectors, in optical communication with the array of VCSELs and the diffractive optical element, to detect interference between respective fanned-out copies of the first beam from the diffractive optical element and the second beams,
wherein the array of photodetectors is configured to generate outputs proportional to
∑
t
=
1
k
[
W
j
i
1
-
(
X
i
)
2
-
X
i
1
-
(
W
j
i
)
2
]
,
where X i represents the i th element of the activation vector and W j i represents the ij th element of the weight matrix.