Digital radio links fail with a "cliff": once channel noise passes the threshold the error-correcting code was designed for, the receiver gets nothing usable. Deep joint source-channel coding (DeepJSCC) is a neural autoencoder that maps an image straight to transmitted signal values and decodes it from the noisy received signal. It avoids the cliff because noise produces gradual blurring instead of total failure. The appeal of DeepJSCC is strongest on small, battery-powered radios, and those are exactly the devices that would run the encoder and decoder on analog in-memory chips. These chips compute matrix products inside arrays of resistive memory cells. They are very energy-efficient, but their stored weights are noisy, drift over time, and pass through coarse converters. Nobody has checked whether "no cliff" still holds when the noise comes from the compute hardware as well as from the air.
This note proposes to measure that directly and to fix it if it breaks. We treat the whole chain as one compound channel: a noisy analog encoder, a frequency-selective multipath radio channel, and a drifting analog decoder. We then ask whether reconstruction quality degrades smoothly across the joint space of channel SNR and hardware age. Our hypothesis has two parts. First, transmitter-side compute noise behaves much like extra channel noise, which DeepJSCC already tolerates. Second, receiver-side weight drift and structured device faults are not channel-like, and can produce a new, hardware-induced cliff that standard noise-injection training does not remove.
Core Claim
Graceful degradation is a property of the noise a code was trained against, not of joint source-channel coding as such. Analog compute noise enters a neural transceiver in places where it is not equivalent to channel noise. These are the receiver's equalization and denoising layers, and structured faults such as stuck cells and wire-resistance gradients. In those places it can create sharp quality drops even when the channel is good.
The fix we propose is location-matched stress training. Hardware noise is injected where it physically occurs, on each analog layer of both encoder and decoder. Its severity is randomized jointly with channel SNR during training. The decoder is also told an estimate of its own hardware state, read from on-chip reference cells. This mirrors how SwinJSCC conditions on channel SNR.
The prediction that could be refuted is specific. Two simpler approaches should each leave a visible cliff along the hardware-age axis: plain hardware-aware training at a fixed noise level, and plain DeepJSCC trained at a lower channel SNR to "absorb" compute noise. The proposed method should remove that cliff.
Where This Sits
Three lines of work each cover two corners of the problem.
DeepJSCC itself, introduced in Deep Joint Source-Channel Coding for Wireless Image Transmission and scaled up with Swin Transformers and SNR-conditioning modules in SwinJSCC, motivates the work. It shows graceful degradation, but only under additive white Gaussian noise (AWGN) or flat fading, and on floating-point GPUs. DeepJSCC-Q handles one hardware constraint, discrete modulation constellations. That constraint concerns the format of the signal, not noise inside the neural computation. OFDM-Guided DeepJSCC supplies our channel model and architecture. OFDM (orthogonal frequency division multiplexing) splits a wideband multipath channel into many narrow subchannels. That paper shows that realistic multipath propagation breaks DeepJSCC models trained on simplified channels unless OFDM, channel estimation, and equalization are built into the network. It is our strongest evidence that simplified noise models fail in deployment.
On the hardware side, the methodology comes from Rasch et al.'s hardware-aware training study, the aihwkit toolkit, and Spoon et al.. Hardware-aware training means injecting phenomenological device noise into the forward pass while updating precise digital copies of the weights. That body of work supplies three facts that shape our design: output-converter noise hurts more than weight noise, global drift compensation works, and mixed analog/digital partitioning (keeping nonlinear operations in INT6 digital logic) is viable. That work never involves a propagation channel, so it cannot ask how compute noise interacts with channel noise.
Two further works constrain us. Lammie et al. document stuck-at faults that are not Gaussian. Can We Trust LLMs on Memristors? shows that some computations degrade sharply under analog noise while perceptual ones tolerate it. That finding supports our hypothesis that the decoder's inverse-problem layers, rather than its perceptual layers, are where a hardware cliff would appear.
To our knowledge from the file, no prior work concatenates analog compute non-idealities with a multipath channel in one end-to-end model.
Research Program
We will build a compound simulator in PyTorch. The convolutional layers of the OFDM-guided DeepJSCC encoder and decoder become aihwkit analog layers using its PCM inference model, which is calibrated on a one-million-device array and includes programming noise, read noise, power-law drift, and converter quantization. Following Spoon et al.'s partitioning, the FFT and IFFT, cyclic prefix, MMSE equalization, power normalization, and nonlinearities stay in low-precision digital logic. The multipath channel and OFDM operations are differentiable fixed layers, as in Yang et al. We deliberately do not solve the energy question of where to place the equalizer. Instead, the partition is a controlled variable: we run one variant in which the learned channel-estimation and equalization refinement subnets are analog and one in which they are digital.
The central measurement is a degradation surface. Write $Q(s,t)$ for the mean PSNR on Kodak at channel SNR $s$ (dB) and time $t$ since the analog weights were programmed. We define a cliff index as the largest drop in $Q$ between adjacent grid points, normalized by the step size along each axis. A gracefully degrading system has a small cliff index along both axes. Digital separation has a large one along $s$.
We will also attribute damage by location. Noise is switched on in the encoder only, the decoder only, and the refinement subnets only. This answers the open question of whether transmitter or receiver drift drives collapse, including how drift in the transmitter interacts with the transmit power constraint.
The method has three ingredients:
- Location-matched injection. Every analog layer of both networks receives its own noise during training.
- Joint stress randomization. Each training batch samples channel SNR and a drift time spanning one second to one simulated year. One year is the horizon over which Rasch et al. report surviving drift.
- Hardware-state conditioning. A small modulation network rescales decoder features from two inputs: channel SNR and a scalar hardware-noise estimate derived from reference devices. The same reference devices already drive global drift compensation, so this adds no new sensing hardware.
Baselines
The naive reference point is the OFDM-guided DeepJSCC model trained in floating point and mapped directly onto the simulated analog hardware, with and without aihwkit's global drift compensation. This shows how large the problem is. Rasch et al. found that direct mapping of WideResNet-16 raises error from 20% to 25%. We have no comparable figure for a transceiver, which is what this experiment produces.
The strongest digital competitor is BPG compression with 5G LDPC coding over OFDM. Yang et al. report that their floating-point OFDM-guided DeepJSCC achieves a 2.5–4 dB SNR gain over this scheme on multipath channels, and SwinJSCC reports a 25–29% BD-rate saving over BPG+LDPC on Kodak. This baseline runs in digital logic, so it suffers the channel cliff but not the hardware cliff. The analog transceiver beats it only if, after one simulated year, it still keeps a positive SNR gain at equal quality across the low-SNR range.
The strongest analog method is standard hardware-aware training applied to the same transceiver with aihwkit, at a single fixed noise configuration and a fixed training SNR band, as in Rasch et al. No reported result found on Lacuna for hardware-aware-trained DeepJSCC on any image benchmark. Its numbers will therefore come from our own runs. It will use the same Kodak/DIV2K data, the same architecture, the same number of training steps, and the same hyperparameter search budget as our method. Our method beats it if it lowers the cliff index along the drift axis and raises the worst-case PSNR over the full $(s,t)$ grid, while matching its PSNR at $t$ near zero within the spread across seeds.
Two ablations isolate the key ingredients. The first keeps location-matched injection and joint randomization but removes hardware-state conditioning, which tests whether the reference-cell signal matters. The second injects all hardware noise as additive Gaussian noise at the channel input instead of on the layers. This is the "compute noise is just channel noise" hypothesis made into a training method.
Why not just…?
Why not just run standard hardware-aware training on OFDM-guided DeepJSCC? This is the obvious combination and it may get most of the way. Rasch et al. show it recovers iso-accuracy on many classifiers. However, their open-problem statement and our own setup point to a weakness when deployment noise exceeds training noise, and a transceiver must operate over a two-dimensional range of conditions, not one. If fixed-level hardware-aware training already gives a flat degradation surface, our method adds nothing, and we will report that. The decisive check is the cliff index along the drift axis at high channel SNR, where hardware noise dominates.
Why not just train DeepJSCC at a lower SNR so it absorbs compute noise? This is the most serious threat. Transmitter-side output noise really does look like channel noise, and we expect this shortcut to work for the encoder. It is unlikely to hold for decoder weight noise, which corrupts the equalization layers that Yang et al. found essential, and which therefore acts multiplicatively and in a way that depends on the signal. The second ablation is exactly this alternative. If it matches our method, we will conclude that a simple "equivalent SNR" rule is enough. That would be a useful negative result in its own right.
Why not just keep the transceiver in INT8 digital logic? This removes the hardware cliff entirely and is the right choice if energy is not binding. The surveyed in-memory computing literature argues that weight movement dominates energy at the edge, and Spoon et al. show mixed analog/digital designs saving 91% of digital-component energy. We do not claim to settle the energy question. The INT8 digital version is reported as an upper bound on quality, so readers can see what the analog energy savings cost in PSNR.
Why not just rely on drift compensation? Global rescaling corrects the mean decay of the weights but not the cell-to-cell spread. Rasch et al. report that models without hardware-aware training fail over time even with compensation. The naive baseline with compensation tests whether this also holds for transceivers.
Proposed Experiments
We train on DIV2K and CIFAR-10. We evaluate PSNR and MS-SSIM on Kodak, with CLIC2021 as an out-of-distribution set. Channels are multipath OFDM with the power-delay settings of Yang et al., plus AWGN for reference. The evaluation grid spans channel SNR from 0 to 20 dB and drift time from one second to one simulated year, with five seeds. Bandwidth ratios follow the setups of Yang et al.
The first decisive check is the localization map. If decoder-side drift, especially in the refinement subnets, produces a larger cliff index than encoder-side drift at equal noise magnitude, the claim that compute noise is not channel-equivalent is supported. If not, the simpler equivalent-SNR approach wins.
The second decisive check is a test of whether simplified hardware noise models fail the way simplified channel models do. We train every method with aihwkit's Gaussian-plus-power-law phenomenology. We then test on non-idealities held out from training: stuck-at-ON/OFF cells at increasing rates, drift exponents larger than in training, one fewer bit of converter resolution, and a structured spatial gradient across the array as a stand-in for wire-resistance (IR) drop.
This second test addresses Gap 4 of the file. It asks whether training against simplified hardware noise produces a deployment cliff, just as Yang et al. found for training against AWGN. A sharp quality drop under held-out structured faults, for every training method, would be the hardware analogue of the cliff effect. It would also show that the next step is calibrated, spatially structured noise models rather than more Gaussian injection.
References
- Deep Joint Source-Channel Coding for Wireless Image Transmission
- SwinJSCC: Taming Swin Transformer for Deep Joint Source-Channel Coding
- DeepJSCC-Q: Constellation Constrained Deep Joint Source-Channel Coding
- OFDM-Guided Deep Joint Source Channel Coding for Wireless Multipath Fading Channels
- Hardware-aware training for large-scale and diverse deep learning inference workloads using in-memory computing-based accelerators
- A flexible and fast PyTorch toolkit for simulating training and inference on analog crossbar arrays
- Toward Software-Equivalent Accuracy on Transformer-Based Deep Neural Networks With Analog Memory Devices
- Design Space Exploration of Dense and Sparse Mapping Schemes for RRAM Architectures
- Can We Trust LLMs on Memristors? Diving into Reasoning Ability under Non-Ideality