Modern base stations serve many users at once. Each user gets its own modulation order. A user close to the tower might send 256-QAM (quadrature amplitude modulation, which packs 8 bits into one of 256 points on a grid), while a user at the cell edge sends 16-QAM. The chip that separates these users, the MIMO detector, still runs all of them through one fixed-point datapath, sized for the hardest case. The number of bits each arithmetic unit needs is set mainly by how close the grid points are, and that spacing differs roughly fourfold between 16-QAM and 256-QAM. So a detector sized for the densest user wastes bits on every other user. Detector hardware today either fixes one precision for everything, or switches the whole chip between a few global modes.
We propose to give each user its own precision, chosen from a closed-form rule that ties bitwidth to that user's constellation spacing and signal quality. Users with similar precision needs are then physically regrouped into lanes of equal width. This keeps the hardware regular, avoiding the scattered memory access and bit-unpacking costs that usually erase the savings of mixed precision. A final ingredient addresses the gap between simulated and measured hardware. We will choose among precision plans using energy and throughput measured on an FPGA, not bit counts, and test whether the proxy and the measurement disagree.
Core Claim
In a coordinate-descent MIMO detector, each coordinate update belongs to one user. That makes the user the natural unit for assigning precision. We claim that per-user precision, packed into a small number of equal-width lanes by reordering users, lowers measured energy per bit relative to the best single precision. Detection quality stays within a fixed small margin of floating point. The saving should grow with how mixed the users' modulation orders are, and it should vanish when all users share one order. That prediction is falsifiable and built into the experiments.
The precision rule comes from requiring quantization noise to stay a small fraction of the noise the detector already has to tolerate. For a uniform quantizer with step size $\Delta_k$ on user $k$'s signal path, the rounding noise power is $\Delta_k^2/12$. We require
$$\frac{\Delta_k^2}{12} \le \kappa \, \min\!\left(\sigma_k^2,\; \tfrac{1}{4} d_{\min}(M_k)^2\right), \qquad d_{\min}(M) = \sqrt{\tfrac{6}{M-1}},$$
Here $\sigma_k^2$ is user $k$'s post-equalization noise variance, $d_{\min}(M_k)$ is the minimum distance between points of a unit-energy $M_k$-QAM constellation, and $\kappa$ is a tolerance set once for the whole system. Taking logarithms, the fractional bits for user $k$ are about $b_k = \lceil \log_2(1/\Delta_k)\rceil$. Two consequences follow from the $1/\sqrt{M}$ scaling of $d_{\min}$. Each fourfold increase in modulation order costs about one more bit. Each 6 dB of extra signal quality costs about one bit when noise, rather than grid spacing, is the binding limit. The rule is cheap enough to evaluate once per coherence block, the stretch of time over which the channel stays roughly constant, from quantities the scheduler already knows.
Grounding in Prior Work
Three existing detectors motivate the proposal by showing the gap from different sides.
A Deep-Unfolding-Optimized Coordinate-Descent Data-Detector ASIC for mmWave Massive MIMO builds a 22 nm Gram-domain block coordinate-descent detector (GBCD) that reaches 256-QAM for 128 antennas and 16 users. It reports 7.1 Gbps at 367 mW and a 4 to 10 dB SNR margin over LMMSE, the standard linear minimum-mean-square-error detector, in loaded channels. Its datapath precision, however, is fixed. We reuse its algorithm structure, including the SINR-sorted per-user coordinate updates and the piecewise-linear approximation of the posterior mean. We change only how precision is assigned.
A Resolution-Adaptive 8 mm2 9.98 Gb/s 39.7 pJ/b 32-Antenna All-Digital Spatial Equalizer for mmWave Massive MU-MIMO in 65nm CMOS shows that switching the whole converter and equalizer between resolution modes buys about 3× energy efficiency. The switch is global, silicon results stop at 16-QAM, and the authors name higher orders as future work. We move the adaptation from the whole chip to individual users, and we evaluate up to 256-QAM.
Automatic Hybrid-Precision Quantization for MIMO Detectors (AHPQ) assigns a separate fractional bitwidth to each internal variable offline, using a reinforcement-learning agent rewarded by simulated error rate. It reports 58.7% lower average bitwidth than uniform quantization and a 2.97× throughput-to-area gain on an approximate-message-passing ASIC. Its precision is static at run time, blind to which constellation each user is sending, and optimized against simulated error rather than measured cost.
A second group of papers constrains the design. Hardware Constraints on Mixed Precision Quantization Speedups, Channel-Wise Mixed-Precision Quantization for Large Language Models and Dustin: A 16-Cores Parallel Ultra-Low-Power Cluster with 2b-to-32b Fully Flexible Bit-Precision and Vector Lockstep Execution Mode all find the same problem. Fine-grained, ragged bitwidths pay for packing, masking and irregular memory access, and those costs can cancel the arithmetic savings. This is why we reject per-variable or per-antenna allocation and restrict precision to a few lane widths.
The method for making mixed precision regular comes from language-model inference. MicroMix: Efficient Mixed-Precision Quantization with Microscaling Formats for Large Language Models sorts channels into precision classes and fuses the reordering into the quantization kernel, so each block of data has one uniform format with a shared scale. We apply the same reorder-then-pack idea to users in a detector, a setting where it has not been tried. The sort key is our analytic precision rule rather than a sensitivity threshold on activations.
The simulation-to-hardware lens comes from Air Learning: a deep reinforcement learning gym for autonomous aerial robot visual navigation and Accurate and stable empirical CPU power modelling for multi- and many-core systems. Both show that designs tuned on simulated cost drift badly on real hardware, and that feeding measured hardware behavior back into the optimization closes most of the gap. We should be honest about where that lesson applies here. A digital fixed-point detector gives bit-identical outputs in a bit-true simulation and on the FPGA, so the gap we target is in cost (energy, achieved clock rate, stalls from reordering), not in detection error. The radio front end stays simulated, as it does in every cited detector paper.
Research Program
We will first build a bit-true fixed-point model of a GBCD-style detector. In this model, the Gram-matrix products and residual updates for user $k$ run at $b_k$ fractional bits, snapped to one of three lane widths. This model will confirm that the analytic rule keeps each user's SNR loss at a target error rate within a fixed margin of floating point, across 16-, 64- and 256-QAM. It will also show where the rule is loose or tight.
Second, we will implement the detector on an FPGA of the Virtex-7 class used by Algorithm and VLSI Design for 1-Bit Data Detection in Massive MIMO-OFDM and Data Detection in Large Multi-Antenna Wireless Systems via Approximate Semidefinite Relaxation. The design will have a small number of fixed-width processing lanes and a per-block user permutation. Within each lane, the permutation is applied to the Gram matrix rows and columns.
Third, we will measure throughput, achieved clock rate and power on the board's supply rails. These are the quantities the simulation-to-hardware literature says proxies get wrong.
Fourth, we will fit a small cost model from those measurements, using the regressor-pruning approach of the CPU power-modelling paper. With it we will compare two ways of choosing lane widths and the tolerance $\kappa$. One minimizes average bitwidth, as AHPQ does. The other minimizes measured energy per bit.
Baselines
The naive reference is uniform worst-case precision. The whole detector runs at the single bitwidth needed for its densest user at the lowest supported SINR, which is how a fixed-precision chip like the GBCD ASIC is sized in practice. It runs on the same FPGA fabric and the same GBCD algorithm, with bitwidth found by the same bit-true sweep. The published GBCD silicon reports 7.1 Gbps, 367 mW and a 4 to 10 dB margin over LMMSE (A Deep-Unfolding-Optimized Coordinate-Descent Data-Detector ASIC for mmWave Massive MIMO). Those figures come from 22 nm silicon and are not directly comparable to an FPGA, so we compare against our own FPGA port. Beating this baseline means lower measured energy per bit at equal SNR loss relative to floating point.
The strongest simple competitor is a per-block global precision switch. Every coherence block, the whole detector drops to the single bitwidth its current densest user needs. This is the resolution-adaptive chip's strategy (A Resolution-Adaptive 8 mm2 9.98 Gb/s 39.7 pJ/b 32-Antenna All-Digital Spatial Equalizer for mmWave Massive MU-MIMO in 65nm CMOS), which reports 39.7 pJ/b at full resolution and about 3× better efficiency in its low-resolution mode, at 16-QAM. It gets the same rule, the same fabric and the same set of lane widths; only the grouping differs. Beating it means a clear energy-per-bit saving on user mixes with several different modulation orders, which is exactly where our method should win.
The strongest existing mixed-precision method is AHPQ (Automatic Hybrid-Precision Quantization for MIMO Detectors). It reports 58.7% lower average bitwidth than uniform quantization, 2.97× throughput-to-area and 7.93 pJ/b on its message-passing ASIC. We will re-run its two-stage procedure on our GBCD detector, with an integer-bit stage from Monte Carlo value ranges and a learned fractional-bit stage. It gets the same channel data and a tuning budget matched to ours in simulator hours, then goes onto the same FPGA. Beating it means lower measured energy per bit, even if AHPQ achieves a lower average bitwidth. That outcome would directly confirm that bitwidth is a poor proxy.
Two ablations isolate the key ingredients. The first keeps per-user precision but drops the reordering, so users sit in their natural slots with ragged widths. It measures how much of any gain comes from hardware alignment rather than from fewer bits. The second keeps everything but selects lane widths and $\kappa$ by average bitwidth instead of measured energy. It isolates the contribution of the hardware-measurement feedback. A floating-point GBCD run provides the error-rate ceiling.
Why not just…?
Why not just size the detector for 256-QAM once and stop? This is what current silicon does, and it is simple and safe. The rule predicts roughly one extra bit per fourfold increase in modulation order across the whole datapath. That cost is paid even when most scheduled users send 16-QAM. Whether the waste matters in energy terms, not just bit counts, is an empirical question. The uniform worst-case baseline answers it directly on measured power.
Why not just switch the whole chip's precision per block, as the resolution-adaptive equalizer does? This is the real threat. When all users share one modulation order and similar SINR, our method reduces to exactly this switch, and we expect no gain. The advantage exists only for heterogeneous user mixes. We will therefore report the saving as a function of a heterogeneity measure: the spread of required bitwidths across scheduled users, sampled from QuaDRiGa drops with SINR-driven link adaptation. If realistic schedules turn out to be nearly homogeneous, the global switch wins on simplicity, and we will say so.
Why not just use AHPQ's per-variable allocation, which already cuts bitwidth by 58.7%? Its savings are measured in bits and area on a custom ASIC. The mixed-precision systems papers (Hardware Constraints on Mixed Precision Quantization Speedups; Channel-Wise Mixed-Precision Quantization for Large Language Models) show that ragged per-variable widths often lose their advantage once packing and memory access are paid on real hardware. AHPQ's allocation is also fixed offline and cannot follow the scheduler's changing modulation choices. The head-to-head comparison on measured energy per bit, together with the no-reordering ablation, will settle this.
Why not just apply quantization-aware training, or its kernel-based variant, to the detector? Deep Unfolding with Kernel-based Quantization in MIMO Detection shows that standard training-time quantization struggles on unfolded detectors, and that its kernel-based alignment recovers near-FP16 error rate at 8 bits on 16×16 MIMO. It does not go below 8 bits, and it does not vary precision by user. It is complementary: it could tune the piecewise-linear posterior-mean approximation inside each lane. We treat it as an optional add-on rather than a competitor.
Proposed Experiments
Channels will be generated with QuaDRiGa in the mmMAGIC urban micro-cell scenario at 60 GHz, the setup used by both mmWave ASICs cited above. We will test 128×16 and loaded 16×16 antenna-by-user configurations. User modulation orders are drawn from SINR-based link adaptation over 16-, 64- and 256-QAM, plus an optional 1024-QAM stress case. Front-end impairments follow the Rapp power-amplifier and few-bit converter models used in MIMO Detection under Hardware Impairments: Data Augmentation With Boosting, which also lets us check whether the analytic rule stays valid when noise is non-Gaussian.
Detection metrics are per-user uncoded bit and symbol error rate, and SNR loss at a target error rate relative to floating point. Hardware metrics are FPGA resource use, achieved clock rate and measured throughput, rail power, and energy per bit.
Four checks decide the outcome. First, the per-user rule must keep every user within the chosen SNR-loss margin; if it fails at 256-QAM under few-bit converters, the premise fails. Second, measured energy-per-bit savings over the per-block global switch should grow with user heterogeneity and approach zero for homogeneous mixes. A flat curve would mean the gain is not coming from the mechanism we claim. Third, the no-reordering ablation should lose much of the gain on measured power even though its bit counts match ours. That would show lane alignment, not bit reduction alone, carries the result. Fourth, the plan chosen by measured cost should differ from, and beat, the plan chosen by average bitwidth. If they coincide, the simulation-to-hardware ingredient adds nothing for digital detectors, and we will report it as a negative result.
References
- A Deep-Unfolding-Optimized Coordinate-Descent Data-Detector ASIC for mmWave Massive MIMO
- A Resolution-Adaptive 8 mm2 9.98 Gb/s 39.7 pJ/b 32-Antenna All-Digital Spatial Equalizer for mmWave Massive MU-MIMO in 65nm CMOS
- Automatic Hybrid-Precision Quantization for MIMO Detectors
- Channel-Wise Mixed-Precision Quantization for Large Language Models
- Dustin: A 16-Cores Parallel Ultra-Low-Power Cluster With 2b-to-32b Fully Flexible Bit-Precision and Vector Lockstep Execution Mode
- MicroMix: Efficient Mixed-Precision Quantization with Microscaling Formats for Large Language Models
- Air Learning: a deep reinforcement learning gym for autonomous aerial robot visual navigation
- Accurate and stable empirical CPU power modelling for multi- and many-core systems
- Algorithm and VLSI Design for 1-Bit Data Detection in Massive MIMO-OFDM
- Data Detection in Large Multi-Antenna Wireless Systems via Approximate Semidefinite Relaxation
- Deep Unfolding with Kernel-based Quantization in MIMO Detection
- MIMO Detection Under Hardware Impairments: Data Augmentation With Boosting