Homodyne Photonic Tensor Core

• Homodyne photonic tensor cores are photonic computing architectures that use coherent interference and balanced detection to extract multiplication from optical signals.
• They feature diverse designs—including TFLN, hybrid TFLN–Si/SiN, and differential interferometric approaches—that optimize precision, throughput, and energy efficiency.
• Integration of electronic calibration and iterative refinement enables these systems to achieve accurate AI inference and scientific simulation despite analog non-idealities.

Homodyne photonic tensor cores are photonic computing architectures in which tensor operations are implemented through coherent interference and balanced detection, so that multiplication is extracted from the homodyne interference term rather than from optical intensity alone. In the recent literature, this category includes mixed-precision optoelectronic matrix-multiply units built on thin-film lithium niobate (TFLN), hybrid TFLN–Si/SiN coherent GEMM engines, spatiotemporally interleaved homodyne crossbars with time-integrating bus readout, stochastic vector dot-product engines with homodyne accumulation, and differential interferometric tensor cores that directly encode signed operands in phase [2602.08269; 2606.16150; 2604.18496; 2604.09759; 2605.23051]. Across these variants, the defining operation is a balanced photocurrent proportional to a product such as $x \cdot w$, $\Re{E_W E_x^*}$, or a differential interferometric approximation to a dot product, after which time integration, digital accumulation, or mixed-precision iterative correction reconstructs matrix–vector or matrix–matrix results.

1. Coherent multiplication and homodyne readout

The fundamental mechanism is coherent mixing of two optical fields followed by balanced photodetection. In the mixed-precision TFLN tensor core, a continuous-wave laser is split into “X” and “W” paths; travelling-wave amplitude modulators encode the magnitudes and phase modulators encode the phases, so that complex multiplication is represented as $E_1(t)=A_x e^{j\phi_x}$ and $E_2(t)=A_w e^{-j\phi_w}$. A balanced $90^\circ$ optical hybrid and balanced photodiodes then yield photocurrents proportional to the real and imaginary parts of the inner product, with $I_{\text{real}} \propto A_x A_w \cos(\phi_x+\phi_w)$ and $I_{\text{imag}} \propto A_x A_w \sin(\phi_x+\phi_w)$; for real-valued MVM the phase path is unused and the operation collapses to $y=\sum_i A_{x,i}A_{w,i}$ [2602.08269].

The same principle appears in the spatiotemporally interleaved HPTC, where one field acts as signal and the other as local oscillator. If $E_{\text{sig}}\approx \sqrt{P_{\text{sig}}}e^{j\phi_{\text{sig}}}$ and $E_{\text{LO}}\approx \sqrt{P_{\text{LO}}}e^{j\phi_{\text{LO}}}$ are combined in a $2\times2$ MMI, the balanced differential current is
$$
\Delta I = |E_{LO}+E_{sig}|^2 - |E_{LO}-E_{sig}|^2
= 2\sqrt{P_{LO}P_{sig}}\cos(\phi_{sig}-\phi_{LO}),
$$
so phase biasing at $0$ or $\pi$ produces a current proportional to $x\cdot w$ [2606.16150].

A related differential-interferometric derivation underlies DUET. There, two signed analog drives are linearly mapped to phase shifts in an MZI, giving output intensities $I_+=I_0\cos^2[(\phi_{in}+\phi_w)/2]$ and $I_-=I_0\cos^2[(\phi_{in}-\phi_w)/2]$, with balanced photocurrent
$$
\Delta I = I_+ - I_- = 2I_0\sin\phi_{in}\sin\phi_w.
$$
For small angles, $\sin\phi\approx\phi$, so $\Delta I \approx 2I_0(\alpha_x x)(\alpha_w w)\propto x\cdot w$; cascading $k$ segments makes the total differential current proportional to a length-$k$ dot product [2605.23051].

ASTRA preserves the same homodyne accumulation logic but combines it with stochastic optical multiplication. After optical AND gating generates unary pulse streams, a balanced homodyne receiver produces
$$
I_{\text{diff}}[n]=2R\,\Re{E_{\text{sig}}[n]E_{LO}^*},
$$
and the integrated charge over the full bit-stream becomes proportional to $\sum_{n=1}^L b_\otimes[n]$, hence to the product of the encoded operands [2604.09759].

2. Architectural organizations

The architecture space is broad, but recent homodyne tensor cores share a common pattern: high-bandwidth optical encoding, local coherent multiplication, and lower-speed electronic accumulation or correction.

The mixed-precision optoelectronic tensor core reported in “Quantization-aware Photonic Homodyne computing for Accelerated Artificial Intelligence and Scientific Simulation” is built from TFLN modulators, high-speed homodyne detectors, per-channel frequency-domain equalization, and low-speed integrate-and-dump readout ADCs. Weight and activation vectors are time-multiplexed into optical pulse sequences, and a host CPU orchestrates waveform generation, equalization filters, data movement, and digital post-processing [2602.08269].

The spatiotemporally interleaved HPTC replaces full $O(n^2)$ interface replication with two coupled subsystems: a homodyne-crossbar photonic matrix and a bus-readout time-integrating array. Data modulators drive horizontal row waveguides, weight modulators drive orthogonal column waveguides, and each intersection contains a local homodyne detector. Temporal reuse of the detector array and charge accumulation on shared buses reduce the high-speed DAC/modulator and ADC overhead from $O(n^2)$ to $O(n)$ [2606.16150].

The reticle-scale GEMM engine in “Tensor Processing with Homodyne Photonic Integrated Circuits exceeds 1,000 TOPS” uses time multiplexing to reduce required modulators from $O(N^2)$ to $O(N)$, enabling a dense $256\times256$ homodyne array. In that system, wafer-scale fabricated 64-channel TFLN transmitters encode data and chip-to-chip couple to Si/SiN computing circuits containing $N\times K$ homodyne interferometers, with balanced photodiodes, TIAs, ADCs, and FPGA real-time post-processing [2604.18496].

ASTRA uses a different frontend: hundreds to thousands of optical stochastic signed multipliers fan into a single homodyne accumulation stage. The scaling analysis in “Scaling Photonic Tensor Cores with Unary and Homodyne Designs” classifies this as an MWA-organized, unary-encoded, single-wavelength homodyne design, notable for decoupling fan-in from multi-wavelength FSR limits while accumulating many channels on one balanced receiver [2604.14664].

DUET is organized around the vectorized operand differential interferometric cell (VODIC), in which signed inputs and weights are directly mapped to phase shifts in cascaded phase-shifting segments. This avoids sign-splitting and nonlinear remapping, and the same cell can be tiled in time or wavelength multiplexing to implement larger matrix operations [2605.23051].

3. Precision, quantization, and algorithm–hardware co-design

A central issue for homodyne photonic tensor cores is that coherent linearity at the physical layer does not by itself guarantee end-to-end numerical precision. The TFLN mixed-precision study states this explicitly: at low rates of $50\,\text{MS/s}$ the raw analog precision reaches up to $9$ bits with $\sigma\approx0.39\%$, but at $128\,\text{GS/s}$ electro-optic distortion in modulators, cables, and detectors degrades the error to approximately $10\%$, or approximately $3.3$ bits. Measuring each channel transfer function $H(\omega)=Y(\omega)/X(\omega)$ and applying $1/H(\omega)$ pre-emphasis reduces the standard deviation to $\sigma\approx1.72\%$, corresponding to approximately $6$–$7$ bits at $128\,\text{GS/s}$ [2602.08269].

That same work couples calibration to mixed-precision numerical methods. The analog MVM is modeled as $\hat{y}=\bar{A}x+\epsilon$ with $|\epsilon|\sim O(2^{-B})$, and iterative refinement separates a higher-precision outer loop from a lower-precision optical inner loop. Sparse–dense decomposition splits an ill-conditioned matrix as $A=S+D$, computes $S\cdot v$ digitally at $16$ bits and $D\cdot v$ optically at $8$ bits, and recombines both on the CPU; bit-slicing decomposes $16$-bit operands into four $8$-bit products processed by the optical core and digitally reweighted [2602.08269].

The reticle-scale homodyne GEMM study reports a similar precision-throughput trade-off at larger spatial scale: $7$-bit accuracy with standard deviation $\approx1.65\%$ on an $8\times8$ mesh at $50\,\text{GSa/s}$, $6$-bit accuracy with standard deviation $\approx1.52\%$ at $100\,\text{GSa/s}$, and $5.5$-bit accuracy with standard deviation $\approx2.26\%$ at $120\,\text{GSa/s}$; on a $256\times256$ mesh, columns measured up to column $98$ show statistical errors of approximately $6$–$10\%$, described as $6$ bits [2604.18496].

ASTRA frames precision statistically rather than through multi-bit analog linearity. If the unary bit-stream length is $L$, the accumulated charge has mean $K L p_x p_w$ and variance $K^2 L p_x p_w(1-p_x p_w)$, so the relative error decreases approximately as $1/\sqrt{L}$. With $L=128$ plus one sign-bit stream, the reported end-to-end accuracy loss is less than $1.2\%$ relative to FP32 on large-scale NLP and vision transformers [2604.09759].

DUET uses hardware-aware training rather than iterative refinement. Nonidealities are modeled by a differentiable surrogate $\hat{y}=s+u(s)+\epsilon$, with $\epsilon\sim\mathcal{N}(0,\sigma^2(s))$, and the training loop inserts this wrapper around each dot product while quantizing $x$ and $w$ to $5$ bits with a learnable scale factor. The reported calibrated operating range is $\phi\in[-1.52,1.52]\,\text{rad}$ with linearity error less than $1.5\%$, and system SNR is approximately $32\,\text{dB}$, corresponding to effective precision of approximately $5.3$ bits [2605.23051].

4. Throughput, latency, energy efficiency, and interface scaling

The performance envelope of homodyne photonic tensor cores is defined jointly by symbol rate, spatial parallelism, interface overhead, and the degree to which optoelectronic conversion can be amortized across many multiply-accumulate sites.

The following representative metrics are reported for recent systems:

System	Stated scale / rate	Stated precision / efficiency
Mixed-precision TFLN core [2602.08269]	$128\,\text{GS/s}$; $\approx6\,\text{ns}$ latency	$6$–$7$ bit optical; $12$-bit-equivalent solver; $3\,\text{TOPS/W}$
Spatiotemporally interleaved HPTC [2606.16150]	$4\times4$ prototype; $37.8\,\text{GHz}$ EO bandwidth	standard-deviation error $0.0621$ at $10\,\text{Mbaud}$
Reticle-scale coherent GEMM [2604.18496]	$1{,}000$–$6{,}000\,\text{TOPS}$; up to $120\,\text{Gbaud/s}$	$6$–$7$ bit; $330\,\text{TOPS/W}$
ASTRA VDPE [2604.09759]	$30\,\text{TOPS}$ per wavelength; $\sim1\,\text{Peta-ops/s}$ with $W=8$, $M=4$	latency $\approx4.3\,\text{ns}$; less than $1.2\%$ model accuracy loss
DUET projections [2605.23051]	$20\,\text{G}\,\text{symbols/s}$ per segment	$6.01\,\text{TOPS/mm}^2$ projected; $9.52\,\text{TOPS/W}$ weight-stationary

For the TFLN mixed-precision core, a vector-length-$M$ MVM at $128\,\text{GS/s}$ requires $T\approx M\cdot7.8\,\text{ps}$, giving $6.125\,\text{ns}$ for $M=784$. One complex homodyne MVM of length $M$ is counted as $6$ real operations at $128\,\text{GS/s}$, or $768\,\text{GOPS}$; a $100\times100$ crossbar is stated to yield approximately $76.8\,\text{TOPS/chip}$, with projected $10^4\times$ scaling above $15\,\text{POPS}$. The reported present energy figure is approximately $520\,\text{mW}$ for one modulator pair, corresponding to $3\,\text{TOPS/W}$, while integrated ODACs and crossbar fanout are projected above $10{,}000\,\text{TOPS/W}$ [2602.08269].

The reticle-scale coherent GEMM engine emphasizes the throughput law $\text{TOPS}\simeq2N^2f$. With $N=256$ and $f=20\,\text{GSa/s}$, the paper gives $T\approx2.26\times10^{12}\,\text{ops/s}=2{,}262\,\text{TOPS}$, and reports $1{,}000$–$6{,}000\,\text{TOPS}$ over $256\times100$ channels at $20$–$128\,\text{Gbaud/s}$. There the key system argument is that massive $N\times K$ parallelism amortizes DAC, TIA, and ADC cost, leading to a stated efficiency of $330\,\text{TOPS/W}$ at approximately $8\,\text{W}$ total power [2604.18496].

The spatiotemporally interleaved HPTC reframes performance in terms of interface complexity. By building $Y_{n\times n}=W_{n\times t}X_{t\times n}$ as a sum of $t$ rank-1 outer products and reusing the same detector array over time, it reduces write-side electro-optic interfaces from $n^2$ to $2n$, readout chains from $n^2$ to $n$, and total write-plus-read hardware from $O(n^2)$ to $O(n)$. The same paper states that eliminating multi-beam optical combining removes the $1/n$ optical loss typical of passive mesh crossbars and changes required input-power scaling from $O(n^3)$ to $O(n^2)$ [2606.16150].

ASTRA and the comparative scaling study make a related but distinct point: with unary encoding and single-wavelength homodyne accumulation, spatial MAC count can remain invariant with data rate because receiver sensitivity is tied to binary on/off signaling rather than analog amplitude precision. Table I of the scaling paper reports $25{,}600$ MAC lanes for ASTRA’s unary-homodyne MWA design at $1$, $5$, and $10\,\text{GS/s}$, versus smaller or rate-collapsing fan-in for the heterodyne and analog alternatives analyzed there [2604.14664].

5. Workloads and empirical demonstrations

Recent homodyne photonic tensor cores have been evaluated on both AI workloads and scientific simulation, and the reported demonstrations cover real-valued, complex-valued, stochastic, and mixed-precision operating regimes.

For AI inference, the mixed-precision TFLN system demonstrated a two-layer complex-valued neural network on MNIST with topology $784\rightarrow12\rightarrow10$ at $100\,\text{MS/s}$, where optical calibration improved classification from $90.1\%$ to $93.4\%$ against a digital reference of $94.8\%$. A single-layer real network with topology $784\rightarrow10$ at $128\,\text{GS/s}$ evaluated one image in $6.125\,\text{ns}$ and reported optical accuracy of $92.16\%$ versus digital $95.10\%$ on $102$ test images [2602.08269].

At larger scale, the reticle-scale coherent GEMM engine benchmarked Qwen2.5-0.5B. The optical processing unit executed the prefill and decode GEMMs of the model, sustained real-time token generation for batch sizes and context lengths typical of LLM workloads at $20\,\text{GSa/s}$, reduced token-generation latency per iteration below $1\,\text{ms}$, and kept model quality within less than $1\%$ of a digital GPU baseline when measured by cross-entropy and next-token accuracy [2604.18496].

DUET extends the workload range beyond standard classification. Reported results include on-chip Fashion-MNIST accuracy of $90.24\%$ versus digital $91.08\%$, GTSRB macro-average accuracy of $91.71\%$ versus digital $93.35\%$, and BraTS U-Net Dice scores of $0.761$ versus $0.772$ for Whole Tumor, $0.671$ versus $0.684$ for Tumor Core, and $0.747$ versus $0.759$ for Enhancing Tumor. The same work also places dynamic self-attention $QK^\top$ and $PV$ on DUET in a $25.8$M-parameter autoregressive language model and reports qualitatively coherent next-token generation on WikiText [2605.23051].

Scientific computing is prominent in the mixed-precision homodyne literature. The TFLN system reported thin-wire electrostatics on a $100\times100$ BIE with mixed-precision PCG converging in $2$ outer iterations and $11$ optical inner MVMs, achieving $\sigma<0.25\%$ charge-density error; a $101\times101$ 1D EM scattering MoM+GMRES problem with sparse–dense splitting and complex homodyne in $200$ inner iterations, reaching $0.2\%$ residual; and a $15{,}800\times15{,}800$ 3D aircraft RCS problem using bit-sliced inner GMRES with four $8$-bit optical MVMs per $16$-bit product plus three outer digital MVMs, reaching final RCS error $4\times10^{-4}$ over $41\,\text{sr}$ [2602.08269].

ASTRA targets transformer inference rather than PDE solvers, and its abstract reports at least $7.6\times$ speedup and $1.3\times$ lower energy overheads compared to state-of-the-art accelerators, positioning stochastic homodyne accumulation as an alternative route to transformer-scale photonic tensor processing [2604.09759].

6. Limitations, trade-offs, and recurring misconceptions

The recent literature converges on several limitations. First, these systems are not purely optical computers in the sense of eliminating electronic control and correction. The mixed-precision TFLN engine depends on a host CPU for waveform generation, equalization, data movement, and digital post-processing, while the reticle-scale GEMM engine relies on TIAs, ADCs, and FPGA post-processing; this suggests that current homodyne photonic tensor cores are best understood as mixed-signal accelerators rather than all-optical replacements for digital processors [2602.08269; 2604.18496].

Second, homodyne detection improves linearity of multiplication but does not remove calibration burdens. The spatiotemporally interleaved HPTC identifies thermal noise on integration capacitors, shot noise in photodiodes, and phase jitter in thermo-optic shifters; the reticle-scale GEMM work points to off-chip modulators, phase drift, packaging losses, and ADC/TIA bandwidth scaling; DUET emphasizes peripheral-electronics overhead, large-scale calibration, thermal cross-talk, insertion loss, and the device-speed-versus-linearity trade-off [2606.16150; 2604.18496; 2605.23051].

Third, different homodyne tensor-core organizations optimize different bottlenecks. The comparative scaling study argues that the unary-homodyne MWA design offers the strongest path to higher parallelism because single-wavelength operation eliminates FSR and inter-wavelength crosstalk caps, and unary encoding makes spatial MAC count invariant with data rate. The same analysis also states the costs clearly: temporal throughput per weight scales with unary bit-stream length, splitter-tree loss grows as $10\log_2 M$, and area and local-oscillator distribution become significant [2604.14664].

A common misconception is that homodyne photonic tensor cores are intrinsically high-precision analog machines. The reported data do not support that simplification. Raw precision ranges from standard-deviation error $0.0621$ at $10\,\text{Mbaud}$ on a $4\times4$ prototype, to approximately $6$–$7$ bits at $128\,\text{GS/s}$ after equalization in TFLN, to approximately $5.3$ effective bits in DUET, to stochastic precision governed by $1/\sqrt{L}$ in ASTRA. In practice, high-fidelity results are obtained through equalization, iterative refinement, sparse–dense decomposition, bit-slicing, hardware-aware training, or stochastic averaging rather than through the optical core alone [2606.16150; 2602.08269; 2605.23051; 2604.09759].

Taken together, these results indicate that the significance of the homodyne photonic tensor core lies less in a single canonical circuit than in a reusable computational primitive: coherent field-product extraction with balanced detection, coupled to architecture-specific strategies for interface reduction, numerical error management, and workload mapping. A plausible implication is that future progress will depend as much on photonic–electronic co-design and calibration methodology as on optical device bandwidth or raw photonic parallelism.