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 (Zhou et al., 9 Feb 2026, Nie et al., 15 Jun 2026, Zhou et al., 20 Apr 2026, Afifi et al., 10 Apr 2026, Ning et al., 21 May 2026). 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 $y=\sum_i A_{x,i}A_{w,i}$ (Zhou et al., 9 Feb 2026).

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 $\Re\{E_W E_x^*\}$0 MMI, the balanced differential current is

$\Re\{E_W E_x^*\}$1

so phase biasing at $\Re\{E_W E_x^*\}$2 or $\Re\{E_W E_x^*\}$3 produces a current proportional to $\Re\{E_W E_x^*\}$4 (Nie et al., 15 Jun 2026).

A related differential-interferometric derivation underlies DUET. There, two signed analog drives are linearly mapped to phase shifts in an MZI, giving output intensities $\Re\{E_W E_x^*\}$5 and $\Re\{E_W E_x^*\}$6, with balanced photocurrent

$\Re\{E_W E_x^*\}$7

For small angles, $\Re\{E_W E_x^*\}$8, so $\Re\{E_W E_x^*\}$9; cascading $E_1(t)=A_x e^{j\phi_x}$0 segments makes the total differential current proportional to a length-$E_1(t)=A_x e^{j\phi_x}$1 dot product (Ning et al., 21 May 2026).

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

$E_1(t)=A_x e^{j\phi_x}$2

and the integrated charge over the full bit-stream becomes proportional to $E_1(t)=A_x e^{j\phi_x}$3, hence to the product of the encoded operands (Afifi et al., 10 Apr 2026).

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 (Zhou et al., 9 Feb 2026).

The spatiotemporally interleaved HPTC replaces full $E_1(t)=A_x e^{j\phi_x}$4 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 $E_1(t)=A_x e^{j\phi_x}$5 to $E_1(t)=A_x e^{j\phi_x}$6 (Nie et al., 15 Jun 2026).

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 $E_1(t)=A_x e^{j\phi_x}$7 to $E_1(t)=A_x e^{j\phi_x}$8, enabling a dense $E_1(t)=A_x e^{j\phi_x}$9 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 $E_2(t)=A_w e^{-j\phi_w}$0 homodyne interferometers, with balanced photodiodes, TIAs, ADCs, and FPGA real-time post-processing (Zhou et al., 20 Apr 2026).

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 (Alo et al., 16 Apr 2026).

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 (Ning et al., 21 May 2026).

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 $E_2(t)=A_w e^{-j\phi_w}$1 the raw analog precision reaches up to $E_2(t)=A_w e^{-j\phi_w}$2 bits with $E_2(t)=A_w e^{-j\phi_w}$3, but at $E_2(t)=A_w e^{-j\phi_w}$4 electro-optic distortion in modulators, cables, and detectors degrades the error to approximately $E_2(t)=A_w e^{-j\phi_w}$5, or approximately $E_2(t)=A_w e^{-j\phi_w}$6 bits. Measuring each channel transfer function $E_2(t)=A_w e^{-j\phi_w}$7 and applying $E_2(t)=A_w e^{-j\phi_w}$8 pre-emphasis reduces the standard deviation to $E_2(t)=A_w e^{-j\phi_w}$9, corresponding to approximately $90^\circ$0–$90^\circ$1 bits at $90^\circ$2 (Zhou et al., 9 Feb 2026).

That same work couples calibration to mixed-precision numerical methods. The analog MVM is modeled as $90^\circ$3 with $90^\circ$4, 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 $90^\circ$5, computes $90^\circ$6 digitally at $90^\circ$7 bits and $90^\circ$8 optically at $90^\circ$9 bits, and recombines both on the CPU; bit-slicing decomposes $$I_{\text{real}} \propto A_x A_w \cos(\phi_x+\phi_w)$$0-bit operands into four $$I_{\text{real}} \propto A_x A_w \cos(\phi_x+\phi_w)$$1-bit products processed by the optical core and digitally reweighted (Zhou et al., 9 Feb 2026).

The reticle-scale homodyne GEMM study reports a similar precision-throughput trade-off at larger spatial scale: $$I_{\text{real}} \propto A_x A_w \cos(\phi_x+\phi_w)$$2-bit accuracy with standard deviation $$I_{\text{real}} \propto A_x A_w \cos(\phi_x+\phi_w)$$3 on an $$I_{\text{real}} \propto A_x A_w \cos(\phi_x+\phi_w)$$4 mesh at $$I_{\text{real}} \propto A_x A_w \cos(\phi_x+\phi_w)$$5, $$I_{\text{real}} \propto A_x A_w \cos(\phi_x+\phi_w)$$6-bit accuracy with standard deviation $$I_{\text{real}} \propto A_x A_w \cos(\phi_x+\phi_w)$$7 at $$I_{\text{real}} \propto A_x A_w \cos(\phi_x+\phi_w)$$8, and $$I_{\text{real}} \propto A_x A_w \cos(\phi_x+\phi_w)$$9-bit accuracy with standard deviation $$I_{\text{imag}} \propto A_x A_w \sin(\phi_x+\phi_w)$$0 at $$I_{\text{imag}} \propto A_x A_w \sin(\phi_x+\phi_w)$$1; on a $$I_{\text{imag}} \propto A_x A_w \sin(\phi_x+\phi_w)$$2 mesh, columns measured up to column $$I_{\text{imag}} \propto A_x A_w \sin(\phi_x+\phi_w)$$3 show statistical errors of approximately $$I_{\text{imag}} \propto A_x A_w \sin(\phi_x+\phi_w)$$4–$$I_{\text{imag}} \propto A_x A_w \sin(\phi_x+\phi_w)$$5, described as $$I_{\text{imag}} \propto A_x A_w \sin(\phi_x+\phi_w)$$6 bits (Zhou et al., 20 Apr 2026).

ASTRA frames precision statistically rather than through multi-bit analog linearity. If the unary bit-stream length is $$I_{\text{imag}} \propto A_x A_w \sin(\phi_x+\phi_w)$$7, the accumulated charge has mean $$I_{\text{imag}} \propto A_x A_w \sin(\phi_x+\phi_w)$$8 and variance $$I_{\text{imag}} \propto A_x A_w \sin(\phi_x+\phi_w)$$9, so the relative error decreases approximately as $y=\sum_i A_{x,i}A_{w,i}$0. With $y=\sum_i A_{x,i}A_{w,i}$1 plus one sign-bit stream, the reported end-to-end accuracy loss is less than $y=\sum_i A_{x,i}A_{w,i}$2 relative to FP32 on large-scale NLP and vision transformers (Afifi et al., 10 Apr 2026).

DUET uses hardware-aware training rather than iterative refinement. Nonidealities are modeled by a differentiable surrogate $y=\sum_i A_{x,i}A_{w,i}$3, with $y=\sum_i A_{x,i}A_{w,i}$4, and the training loop inserts this wrapper around each dot product while quantizing $y=\sum_i A_{x,i}A_{w,i}$5 and $y=\sum_i A_{x,i}A_{w,i}$6 to $y=\sum_i A_{x,i}A_{w,i}$7 bits with a learnable scale factor. The reported calibrated operating range is $y=\sum_i A_{x,i}A_{w,i}$8 with linearity error less than $y=\sum_i A_{x,i}A_{w,i}$9, and system SNR is approximately $$E_{\text{sig}}\approx \sqrt{P_{\text{sig}}}e^{j\phi_{\text{sig}}}$$0, corresponding to effective precision of approximately $$E_{\text{sig}}\approx \sqrt{P_{\text{sig}}}e^{j\phi_{\text{sig}}}$$1 bits (Ning et al., 21 May 2026).

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 (Zhou et al., 9 Feb 2026)	$$E_{\text{sig}}\approx \sqrt{P_{\text{sig}}}e^{j\phi_{\text{sig}}}$$2; $$E_{\text{sig}}\approx \sqrt{P_{\text{sig}}}e^{j\phi_{\text{sig}}}$$3 latency	$$E_{\text{sig}}\approx \sqrt{P_{\text{sig}}}e^{j\phi_{\text{sig}}}$$4–$$E_{\text{sig}}\approx \sqrt{P_{\text{sig}}}e^{j\phi_{\text{sig}}}$$5 bit optical; $$E_{\text{sig}}\approx \sqrt{P_{\text{sig}}}e^{j\phi_{\text{sig}}}$$6-bit-equivalent solver; $$E_{\text{sig}}\approx \sqrt{P_{\text{sig}}}e^{j\phi_{\text{sig}}}$$7
Spatiotemporally interleaved HPTC (Nie et al., 15 Jun 2026)	$$E_{\text{sig}}\approx \sqrt{P_{\text{sig}}}e^{j\phi_{\text{sig}}}$$8 prototype; $$E_{\text{sig}}\approx \sqrt{P_{\text{sig}}}e^{j\phi_{\text{sig}}}$$9 EO bandwidth	standard-deviation error $$E_{\text{LO}}\approx \sqrt{P_{\text{LO}}}e^{j\phi_{\text{LO}}}$$0 at $$E_{\text{LO}}\approx \sqrt{P_{\text{LO}}}e^{j\phi_{\text{LO}}}$$1
Reticle-scale coherent GEMM (Zhou et al., 20 Apr 2026)	$$E_{\text{LO}}\approx \sqrt{P_{\text{LO}}}e^{j\phi_{\text{LO}}}$$2–$$E_{\text{LO}}\approx \sqrt{P_{\text{LO}}}e^{j\phi_{\text{LO}}}$$3; up to $$E_{\text{LO}}\approx \sqrt{P_{\text{LO}}}e^{j\phi_{\text{LO}}}$$4	$$E_{\text{LO}}\approx \sqrt{P_{\text{LO}}}e^{j\phi_{\text{LO}}}$$5–$$E_{\text{LO}}\approx \sqrt{P_{\text{LO}}}e^{j\phi_{\text{LO}}}$$6 bit; $$E_{\text{LO}}\approx \sqrt{P_{\text{LO}}}e^{j\phi_{\text{LO}}}$$7
ASTRA VDPE (Afifi et al., 10 Apr 2026)	$$E_{\text{LO}}\approx \sqrt{P_{\text{LO}}}e^{j\phi_{\text{LO}}}$$8 per wavelength; $$E_{\text{LO}}\approx \sqrt{P_{\text{LO}}}e^{j\phi_{\text{LO}}}$$9 with $\Re\{E_W E_x^*\}$00, $\Re\{E_W E_x^*\}$01	latency $\Re\{E_W E_x^*\}$02; less than $\Re\{E_W E_x^*\}$03 model accuracy loss
DUET projections (Ning et al., 21 May 2026)	$\Re\{E_W E_x^*\}$04 per segment	$\Re\{E_W E_x^*\}$05 projected; $\Re\{E_W E_x^*\}$06 weight-stationary

For the TFLN mixed-precision core, a vector-length-$\Re\{E_W E_x^*\}$07 MVM at $\Re\{E_W E_x^*\}$08 requires $\Re\{E_W E_x^*\}$09, giving $\Re\{E_W E_x^*\}$10 for $\Re\{E_W E_x^*\}$11. One complex homodyne MVM of length $\Re\{E_W E_x^*\}$12 is counted as $\Re\{E_W E_x^*\}$13 real operations at $\Re\{E_W E_x^*\}$14, or $\Re\{E_W E_x^*\}$15; a $\Re\{E_W E_x^*\}$16 crossbar is stated to yield approximately $\Re\{E_W E_x^*\}$17, with projected $\Re\{E_W E_x^*\}$18 scaling above $\Re\{E_W E_x^*\}$19. The reported present energy figure is approximately $\Re\{E_W E_x^*\}$20 for one modulator pair, corresponding to $\Re\{E_W E_x^*\}$21, while integrated ODACs and crossbar fanout are projected above $\Re\{E_W E_x^*\}$22 (Zhou et al., 9 Feb 2026).

The reticle-scale coherent GEMM engine emphasizes the throughput law $\Re\{E_W E_x^*\}$23. With $\Re\{E_W E_x^*\}$24 and $\Re\{E_W E_x^*\}$25, the paper gives $\Re\{E_W E_x^*\}$26, and reports $\Re\{E_W E_x^*\}$27–$\Re\{E_W E_x^*\}$28 over $\Re\{E_W E_x^*\}$29 channels at $\Re\{E_W E_x^*\}$30–$\Re\{E_W E_x^*\}$31. There the key system argument is that massive $\Re\{E_W E_x^*\}$32 parallelism amortizes DAC, TIA, and ADC cost, leading to a stated efficiency of $\Re\{E_W E_x^*\}$33 at approximately $\Re\{E_W E_x^*\}$34 total power (Zhou et al., 20 Apr 2026).

The spatiotemporally interleaved HPTC reframes performance in terms of interface complexity. By building $\Re\{E_W E_x^*\}$35 as a sum of $\Re\{E_W E_x^*\}$36 rank-1 outer products and reusing the same detector array over time, it reduces write-side electro-optic interfaces from $\Re\{E_W E_x^*\}$37 to $\Re\{E_W E_x^*\}$38, readout chains from $\Re\{E_W E_x^*\}$39 to $\Re\{E_W E_x^*\}$40, and total write-plus-read hardware from $\Re\{E_W E_x^*\}$41 to $\Re\{E_W E_x^*\}$42. The same paper states that eliminating multi-beam optical combining removes the $\Re\{E_W E_x^*\}$43 optical loss typical of passive mesh crossbars and changes required input-power scaling from $\Re\{E_W E_x^*\}$44 to $\Re\{E_W E_x^*\}$45 (Nie et al., 15 Jun 2026).

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 $\Re\{E_W E_x^*\}$46 MAC lanes for ASTRA’s unary-homodyne MWA design at $\Re\{E_W E_x^*\}$47, $\Re\{E_W E_x^*\}$48, and $\Re\{E_W E_x^*\}$49, versus smaller or rate-collapsing fan-in for the heterodyne and analog alternatives analyzed there (Alo et al., 16 Apr 2026).

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 $\Re\{E_W E_x^*\}$50 at $\Re\{E_W E_x^*\}$51, where optical calibration improved classification from $\Re\{E_W E_x^*\}$52 to $\Re\{E_W E_x^*\}$53 against a digital reference of $\Re\{E_W E_x^*\}$54. A single-layer real network with topology $\Re\{E_W E_x^*\}$55 at $\Re\{E_W E_x^*\}$56 evaluated one image in $\Re\{E_W E_x^*\}$57 and reported optical accuracy of $\Re\{E_W E_x^*\}$58 versus digital $\Re\{E_W E_x^*\}$59 on $\Re\{E_W E_x^*\}$60 test images (Zhou et al., 9 Feb 2026).

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 $\Re\{E_W E_x^*\}$61, reduced token-generation latency per iteration below $\Re\{E_W E_x^*\}$62, and kept model quality within less than $\Re\{E_W E_x^*\}$63 of a digital GPU baseline when measured by cross-entropy and next-token accuracy (Zhou et al., 20 Apr 2026).

DUET extends the workload range beyond standard classification. Reported results include on-chip Fashion-MNIST accuracy of $\Re\{E_W E_x^*\}$64 versus digital $\Re\{E_W E_x^*\}$65, GTSRB macro-average accuracy of $\Re\{E_W E_x^*\}$66 versus digital $\Re\{E_W E_x^*\}$67, and BraTS U-Net Dice scores of $\Re\{E_W E_x^*\}$68 versus $\Re\{E_W E_x^*\}$69 for Whole Tumor, $\Re\{E_W E_x^*\}$70 versus $\Re\{E_W E_x^*\}$71 for Tumor Core, and $\Re\{E_W E_x^*\}$72 versus $\Re\{E_W E_x^*\}$73 for Enhancing Tumor. The same work also places dynamic self-attention $\Re\{E_W E_x^*\}$74 and $\Re\{E_W E_x^*\}$75 on DUET in a $\Re\{E_W E_x^*\}$76M-parameter autoregressive LLM and reports qualitatively coherent next-token generation on WikiText (Ning et al., 21 May 2026).

Scientific computing is prominent in the mixed-precision homodyne literature. The TFLN system reported thin-wire electrostatics on a $\Re\{E_W E_x^*\}$77 BIE with mixed-precision PCG converging in $\Re\{E_W E_x^*\}$78 outer iterations and $\Re\{E_W E_x^*\}$79 optical inner MVMs, achieving $\Re\{E_W E_x^*\}$80 charge-density error; a $\Re\{E_W E_x^*\}$81 1D EM scattering MoM+GMRES problem with sparse–dense splitting and complex homodyne in $\Re\{E_W E_x^*\}$82 inner iterations, reaching $\Re\{E_W E_x^*\}$83 residual; and a $\Re\{E_W E_x^*\}$84 3D aircraft RCS problem using bit-sliced inner GMRES with four $\Re\{E_W E_x^*\}$85-bit optical MVMs per $\Re\{E_W E_x^*\}$86-bit product plus three outer digital MVMs, reaching final RCS error $\Re\{E_W E_x^*\}$87 over $\Re\{E_W E_x^*\}$88 (Zhou et al., 9 Feb 2026).

ASTRA targets transformer inference rather than PDE solvers, and its abstract reports at least $\Re\{E_W E_x^*\}$89 speedup and $\Re\{E_W E_x^*\}$90 lower energy overheads compared to state-of-the-art accelerators, positioning stochastic homodyne accumulation as an alternative route to transformer-scale photonic tensor processing (Afifi et al., 10 Apr 2026).

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 (Zhou et al., 9 Feb 2026, Zhou et al., 20 Apr 2026).

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 (Nie et al., 15 Jun 2026, Zhou et al., 20 Apr 2026, Ning et al., 21 May 2026).

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 $\Re\{E_W E_x^*\}$91, and area and local-oscillator distribution become significant (Alo et al., 16 Apr 2026).

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 $\Re\{E_W E_x^*\}$92 at $\Re\{E_W E_x^*\}$93 on a $\Re\{E_W E_x^*\}$94 prototype, to approximately $\Re\{E_W E_x^*\}$95–$\Re\{E_W E_x^*\}$96 bits at $\Re\{E_W E_x^*\}$97 after equalization in TFLN, to approximately $\Re\{E_W E_x^*\}$98 effective bits in DUET, to stochastic precision governed by $\Re\{E_W E_x^*\}$99 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 (Nie et al., 15 Jun 2026, Zhou et al., 9 Feb 2026, Ning et al., 21 May 2026, Afifi et al., 10 Apr 2026).

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.