PaperlayerView exact revision on arXiv

Source-linked AI summary

AIR-LLM: Broadcasting AI Weights over Radio for Memory-Free Edge LLM Inference via RF Computing

Zhihui Gao, Tingjun Chen, Dirk Englund

arXiv:2610.00465v1cs.ITcs.ETcs.LGeess.SPmath.ITphysics.app-ph

TL;DR

Edge LLM inference is constrained by weight storage and loading energy. AIR-LLM broadcasts weights for RF-domain GEMV and calibrates each user's channel at the edge. Its evaluation reports a small perplexity degradation alongside lower energy and airtime than the stated baselines.

  • Problem

    Edge devices may lack memory for LLM weights, and loading weights from memory to the computing unit can consume unaffordable energy during memory-bound inference.

    Source: §1 · p. 1

  • Method

    AIR-LLM broadcasts OFDM-encoded LLM weights for edge RF mixers to compute GEMV, using MIMO spatial multiplexing and edge precoder–postcoder channel calibration.

    Source: §1 · pp. 1–2

  • Results

    4.0% WikiText-2 perplexity degradation on LLaMA-3.1-8B accompanies 157.7×/40.4× lower per-token energy than FP16/4-bit remote-storage baselines and 104.1×/26.0× shorter airtime with 20 users.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Because weight access is read-only, one user-unaware broadcast can serve arbitrarily many edge users within coverage at constant airtime.

    Source: §1 · p. 1

    The channel-dependent calibration objective assumes zero-mean, mutually independent, identically distributed weights and inputs, and a high-SNR region.

    Source: §3.3 · p. 5

Abstract

from arXiv · show

Next-generation large language models (LLMs) are expanding from the cloud to ubiquitous edge devices. However, edge devices typically either lack the memory to store increasingly large LLM weights or, even with enough memory, spend unaffordable energy on loading the weights. This raises our question: can an edge device run an LLM without storing or loading its weights, but receive them over the air and consume them on the fly? Inspired by wireless broadcasting, we present AIR-LLM, an LLM inference architecture for edge devices, which is composed of: (i) a central radio (e.g., 5G base stations) that broadcasts the LLM weights into the air, and (ii) the edge user that receives the weights and completes the general matrix-vector multiplication (GEMV) of LLM inference directly in the radio frequency (RF) domain using RF mixers. To further shorten the airtime, AIR-LLM exploits MIMO spatial multiplexing and proposes an energy-efficient precoder-postcoder pair on the edge to calibrate its own wireless channel. Since the central radio stays user-unaware, AIR-LLM is user-scalable so that one broadcast serves unlimited users within its coverage. We implement AIR-LLM on the NVIDIA Sionna ray-traced channels of two real-world urban scenes and the profiling of a real RF mixer. With a WikiText-2 perplexity degradation of 4.0% on LLaMA-3.1-8B, AIR-LLM saves the energy by 157.7x/40.4x against the FP16 and weight-only quantization baselines; with 20 users, its airtime is 104.1x/26.0x shorter, respectively.

1 Introduction

AIR-LLM addresses edge LLM memory and weight-loading energy constraints by broadcasting weights for RF-domain inference, while avoiding user-specific transmission setup. Its evaluations report near-lossless perplexity with substantially lower energy and airtime than the stated baselines.

1 Introduction

  • Edge devices may lack memory for LLM weights, and even devices that can store them face unaffordable energy costs from loading weights during memory-bound inference.LLaMA-3.1-8B occupies 16 GB in FP16 and 4 GB after 4-bit weight-only quantization.Source: §1 · p. 1
  • Prior in-memory and over-the-air computing approaches have GEMV sizes limited by hardware complexity, which is incompatible with growing LLM GEMVs.The paper cites crossbars and antenna arrays as examples of hardware-limited sizes.Source: §1 · p. 1
  • One broadcast suits read-only weight access, and the central radio remains user-unaware; AIR-LLM also needs no handshake or per-user modulation and coding scheme.The waveform stays identical regardless of users’ SNRs, and one transmission serves every user within coverage.Source: §1 · pp. 1–2
  • AIR-LLM broadcasts OFDM-encoded weights for edge RF mixers to compute GEMV outputs, while MIMO spatial multiplexing shortens airtime and edge calibration compensates for each user's channel.A calibration unit pairs a per-subcarrier precoder with a per-antenna postcoder.Source: §1 · p. 2
  • 4.0% WikiText-2 perplexity degradation on LLaMA-3.1-8B accompanies 157.7×/40.4× lower per-token energy than FP16/4-bit remote-storage baselines and 104.1×/26.0× shorter airtime with 20 users.Energy is also 51.5×/13.9× lower than fetching FP16/4-bit weights from local memory.Source: §1 · p. 2

2 Preliminaries

The preliminaries describe the OFDM/MIMO channel representation and explain how an RF mixer’s linear convolution can produce GEMV inner products. For real-valued GEMVs, paired inner products are recovered from one output subcarrier.

2.1 OFDM and MIMO Channel Model

  • The OFDM/MIMO model represents transmitted and received frequency-domain signals across antennas and subcarriers, with a channel response for each antenna pair and subcarrier.Conventional edge-side channel recovery requires 8A_tA_rN_fft FLOPs.Source: §2.1 · pp. 2–3

2.2 Analog GEMV via RF Mixers

  • An RF mixer performs complex linear convolution, whose middle output subcarrier yields two real-valued inner products from the real and imaginary parts of a paired weight signal.The paired weights encode two rows in reversed subcarrier order; remaining output subcarriers can be discarded after low-pass filtering.Source: §2.2 · p. 3

3 RF Computing Engine

AIR-LLM uses user-unaware MIMO weight broadcasts and edge-side calibration units to recover GEMV outputs despite channel effects. It optimizes the precoder–postcoder pairs using channel-based initialization and least-squares refinement, balancing accuracy, airtime, and energy.

3.1 Problem Formulation

  • An A×A MIMO system supports up to 2A real-valued inner products per round, while the user-unaware broadcast leaves channel effects for edge-side calibration.The conventional per-subcarrier channel inversion is infeasible after filtering mixes subcarriers and would cost 8A^2N_fft FLOPs.Source: §3.1 · pp. 3–4, Figure 2 · p. 4

3.2 Edge Calibration Unit Design

  • Each calibration unit digitally precodes the input, duplicates it across A RF mixers for analog convolutions with received weights, and postcodes the outputs to correct inter-antenna interference.The precoder calibrates frequency response; the postcoder operates across the antenna dimension.Source: §3.2 · p. 4

3.3 Precoder and Postcoder Optimization

  • The original joint precoder–postcoder objective needs weight and input statistics, so AIR-LLM instead uses a channel-dependent objective under zero-mean, independent, identically distributed inputs and weights and a high-SNR assumption.A closed-form SVD approximation initializes the solution, followed by alternating least-squares updates with closed-form steps that typically converge within a handful of iterations.Source: §3.3 · p. 5

3.4 System Efficiency Analysis

  • Airtime decreases with antenna count and bandwidth and is independent of SNR; one broadcast serves arbitrarily many users, while calibration-unit count trades accuracy against energy.Larger C reduces channel-calibration residual but increases energy linearly; larger A shortens airtime but requires more C for accuracy.Source: §3.4 · p. 6

4 Implementation

AIR-LLM is implemented using ray-traced channels from two contrasting urban scenes and profiling of a real RF mixer. The evaluation models wireless propagation and measures mixer distortion separately.

4 Implementation

  • AIR-LLM is evaluated on ray-traced urban channels and with distortion measured from a real RF mixer.The implementation uses two real-world scenes and a wired mixer testbed.Source: §4 · p. 6, §4.1 · p. 6, §4.2 · p. 7

4.1 Channel Modeling by Ray Tracing

  • Paris provides an open, consistently LOS scene, while Munich is densely built with often-blocked LOS and richer multipath.Sionna models propagation by tracing 107 rays through digital city scenes to construct each channel.Source: §4.1 · p. 6

4.2 Hardware Profiling on Real RF Mixer

  • The wired ZEM-4300+ mixer profile isolates mixer distortion and selects the best ENOB over swept LO-port power for each RF-port power.The measured ENOB curve is applied at each edge device’s received weight-signal power.Source: §4.2 · p. 7

5 Evaluation

The evaluation measures GEMV accuracy and energy, LLM perplexity, and airtime across two urban scenes, then examines system settings and multi-user broadcasting. AIR-LLM’s broadcast airtime stays fixed as users increase, while unicasting airtime grows with the number served.

5.1 Benchmarking on GEMV

  • At A=4, raising calibration units from 1 to 8 lifts GEMV ENOB from 0.55 to 3.97 bit; five more units add 0.43 bit at 1.5× energy.Increasing A shortens airtime but requires more calibration and energy for the same ENOB.Source: §5.1 · pp. 7–8
  • As GEMVs scale, AIR-LLM ENOB remains 3.99–4.51 bit in Paris and 3.10–3.39 bit in Munich once multiple rounds are needed.AIR-LLM’s energy per FLOP falls with GEMV size, reaching 50.2× greater energy efficiency than FP16 at N=2^12.Source: Figure 7 · p. 7, §5.1 · p. 8

5.2 Benchmarking on LLM

  • Paris AIR-LLM beats the best WoQ baseline on 7/9 LLMs; Munich beats the best W3A16 baseline on 7/9 LLMs.In Paris it also beats the best analog baseline on 4/9 models and stays within 5.6% of it on the other five.Source: §5.2 · p. 8, Table 2 · p. 8
  • 157.7×/40.4× less per-token energy than remote FP16/W4A16 on LLaMA-3.1-8B; AIR-LLM’s energy advantage holds across all nine LLMs.Against local FP16/W4A16, it uses 51.5×/13.9× less energy, respectively.Source: §5.2 · p. 9, Figure 8 · p. 9

5.3 Impact of the System Factors

  • Paris is saturated by 35 dBm, while Munich remains noise-limited below 50 dBm; at 28 GHz Munich PPL stays above 10^2, unlike the comparable 0.915 GHz results.At 60 dBm, Munich improves to 8.8% above the FP16 reference; 0.915 GHz PPL differs by at most 0.42 across eight (A,C) settings.Source: §5.3 · pp. 9–10, Table 6 · p. 10
  • About A^2/2 calibration units achieve good LLM PPL—roughly half the energy of the A^2-unit theoretical inversion—in the tested channel.For A=4, PPL falls from above 10^5 at two units to 6.49/7.37 in Paris/Munich at eight units, then saturates.Source: §5.3 · p. 10, Table 5 · p. 10

5.4 Broadcasting to Massive Edge Users

  • With 20 users, median PPL is 6.52 in Paris versus W4A16’s 6.65, and 7.23 in Munich versus W3A16’s 8.24.70% of Paris users and 80% of Munich users fall below those respective baselines; one Paris user has PPL 255.27.Source: §5.4 · p. 10, Figure 11 · p. 11
  • 104.1×/26.0× less airtime than FP16/W4A16 with 20 users, while AIR-LLM’s total airtime stays constant as users increase.Unicast baselines serve each user separately, so their total airtime grows linearly with user count.Source: §5.4 · p. 11

6 Related Works

Related work addresses weight movement through quantization, memory-near compute, analog computing, and remote serving. AIR-LLM differs by consuming broadcast weights in the RF domain without storing them on the edge device.

Related work

  • Weight-quantization and memory-near accelerators optimize fetching resident weights; AIR-LLM removes local weight storage and computes on arriving broadcast waveforms.The paper also contrasts its RF mixer’s product operation with over-the-air computing’s summation of multiple transmitters.Source: §6 · p. 11
  • Analog ML accelerators compute physically but face analog-error and converter limits; AIR-LLM reports accuracy in the effective bits resolved by its link.The paper situates AIR-LLM alongside computation embedded in operand delivery over fiber or air.Source: §6 · p. 11
  • Edge-serving approaches either keep fewer weights resident or move the model off-device; AIR-LLM keeps computation and prompts at the edge while broadcasting weights from the radio.One transmission serves devices in coverage, amortizing weight delivery across them.Source: §6 · p. 11

7 Conclusion

AIR-LLM’s appendix extends its RF-computing design with blocked correlation and a channel-calibration procedure. The conclusion characterizes the architecture as memory-access-free, user-scalable, and energy-efficient.

7 Conclusion

  • AIR-LLM is presented as memory-access-free, user-scalable, and energy-efficient, with MIMO shortening airtime and edge precoder–postcoder calibration compensating for the wireless channel.Evaluations on ray-traced channels and a profiled RF mixer show orders-of-magnitude reductions in per-token energy and airtime, with a slight PPL gap to FP16.Source: §7 · p. 11

A Table of Notations

  • The appendix notation tables separately summarize notation from the main body and symbols used only in the appendices.The BCO objective remains bilinear in W and x; under zero-mean i.i.d. entries and high SNR, its expectation reduces to a channel-dependent objective.Source: Table A1 · p. 15, Table A2 · p. 15, §B.2 · p. 16

B Blocked Correlation Operation

  • Appendix B generalizes linear-convolution mapping to blocked correlation, where an RF mixer retains a whole output block of subcarriers.The weights and inputs are correlated in parallel across D blocks, and their outputs accumulate into one output block.Source: B Blocked Correlation Operation · p. 16, §B.1 · p. 16

B.1 Definition and Realization

  • For block length L = L_w + L_x − 1, zero-padding prevents correlations from leaking between blocks, while the middle subcarrier slot accumulates their outputs.Extracting that slot yields the BCO output and reduces the digitized bandwidth and ADC samples by a factor of D.Source: §B.1 · p. 16

B.2 Pre-/Post-coder Optimization

  • Under BCO, the precoder is constant across each block’s L subcarriers, while the postcoder remains shared across subcarriers.The optimization averages channels within each block for a closed-form initialization, then uses alternating closed-form least-squares updates to refine it.Source: §B.1 · p. 16, §B.2 · pp. 16–17

B.3 Convergence of the Optimization

The optimization’s closed-form initialization and ALS refinement recover most channel-compensation accuracy in the first ALS pass. The appendix also analyzes airtime per FLOP for blocked correlation and its dependence on block layout.

B.3 Convergence of the Optimization

  • The closed-form initialization cuts loss from 0.998 to 0.013 in Paris and from 0.947 to 0.259 in Munich, reaching 3.12 and 0.97 bit.Munich’s richer multipath varies within a block, making its averaged per-block channel less representative of the per-subcarrier channel.Source: §B.3 · pp. 17–18
  • 3.79 bit in Paris and 2.85 bit in Munich follow one ALS pass; convergence reaches 4.08 and 3.13 bit, respectively.Iterations 2–10 add only 0.13 bit in Paris and 0.29 bit in Munich; optimization runs once per channel coherence time.Source: §B.3 · p. 18

C Airtime and Energy Consumption

  • The appendix derives airtime and energy metrics first for general BCO blocks, then specializes them to GEMV.Airtime measures how long the edge RF chain is occupied and depends on signal duration set by bandwidth.Source: C Airtime and Energy Consumption · p. 18, §C.1 · p. 18

C.1 Airtime of a BCO

  • Blocked-correlation airtime per FLOP is independent of block count D, and favors layouts with short output blocks relative to the product of input block sizes.Such layouts reuse each transmitted weight across more correlation taps.Source: §C.1 · p. 18

C.2 Energy Consumption of a BCO

The appendix analyzes AIR-LLM’s energy accounting and scaling for BCO and GEMV, then extends blocked correlation to Conv2D. BCO lets convolution reuse broadcast kernels across image positions, with airtime per useful FLOP up to K^2 times lower than GEMV.

C.2 Energy Consumption of a BCO

  • BCO module energy sums seven signal-chain components, with precoding and IFFT paid once per module and conversion, decoding, and summation accumulated across rounds.The ADC term counts L output samples rather than all N_fft samples because the BCO output occupies reduced bandwidth.Source: §C.2 · p. 18

C.3 Airtime of a GEMV

  • Under GEMV mapping, per-FLOP airtime depends on bandwidth and antenna count, not GEMV dimensions or SNR.Both N and M cancel from the scaling law.Source: §C.3 · p. 19

C.4 Energy Consumption of a GEMV

  • GEMV energy per FLOP is amortized over input size, output size, or antenna count, with DAC energy the bottleneck when A is much smaller than N and M.ADC, postcoding, and summation scale as 1/N; precoding and IFFT as 1/M; DAC as 1/A.Source: §C.4 · p. 19

D AIR-LLM in CNN

  • Standard Conv2D does not fit the one-subcarrier-per-output mapping because reusing its kernel across I^2 image positions would require retransmitting it per pixel.Appendix D uses BCO to map Conv2D and analyze its airtime and energy.Source: D AIR-LLM in CNN · p. 19

D.1 Mapping Conv2D onto BCO

  • The Conv2D mapping treats input channels as blocks and image pixels as subcarriers, padding rows to prevent correlation from leaking between them.A kernel is broadcast once for the whole image, while padding adds nominal FLOPs beyond useful convolution FLOPs.Source: §D.1 · pp. 19–20

D.2 Airtime of Conv2D

  • Conv2D airtime per useful FLOP accounts for padding overhead when compared with GEMV airtime.The comparison uses the shared proportionality constant for the two airtime measures.Source: §D.2 · p. 20
  • Conv2D occupies up to K^2 times less airtime per useful FLOP than GEMV, despite padding inefficiency.A single OFDM symbol carrying K^2 kernel taps produces I^2 output pixels through block-wise correlation.Source: §D.2 · p. 20

D.3 Energy Consumption of Conv2D

The Conv2D energy analysis weighs analog savings against digital processing costs, while the evaluation setup characterizes mixer distortion, urban path loss, and comparison baselines.

D.3 Energy Consumption of Conv2D

  • Table A5 organizes one Conv2D layer’s energy complexity and energy per useful FLOP across seven signal-chain terms, assuming I≫K.The analysis substitutes the Conv2D mapping into the per-module energy terms before dividing by useful FLOPs.Source: Table A5 · p. 20, §D.3 · p. 20
  • Conv2D’s ADC and cross-unit-summation energy terms vanish as 1/N with depth, while digital IFFT/FFT costs retain image-size logarithmic factors.Choosing an operating point must weigh converter savings against the digital cost of wider OFDM symbols.Source: §D.3 · pp. 20–21

E Experimental Setup

  • The evaluation setup combines ray-traced channels for two urban scenes with measurements of RF-mixer distortion and each scene’s path-loss distribution.The measurements support the implementation’s characterization of the mixer and candidate-device channels.Source: E Experimental Setup · p. 21

E.1 Measured Distortion and Path Loss

  • Paris path loss is concentrated (median 84.0 dB; 5/95th percentiles 74.9/97.8), while Munich is broader and worse (91.4; 75.2/131.3), spanning noise- to distortion-limited links.The open Paris scene is mostly line-of-sight; Munich has a long tail of devices with blocked line-of-sight.Source: §E.1 · p. 21

E.2 Baselines

  • AIR-LLM is benchmarked across weight precision and storage location, with remote-storage baselines using the same wireless channel and AIR-LLM targeting remote storage.Baselines include FP16, 4/3-bit weight-only quantization, and analog AWGN and HyFlexPIM at AIR-LLM’s ENOB.Source: §E.2 · p. 21

E.3 Evaluation Tasks and Metrics

The evaluation measures GEMV accuracy and whole-LLM quality, alongside communication airtime and energy; a Munich sweep compares ENOB with energy.

E.3 Evaluation Tasks and Metrics

  • The GEMV test draws weights and inputs from i.i.d. Gaussian distributions and quantifies computing accuracy using ENOB.The supplied passage identifies ENOB as the accuracy metric but does not include its full definition.Source: §E.3 · p. 21
  • Figure A4 plots ENOB against energy and shows the minimum energy required for 2–4 ENOB in Munich.The figure repeats the Fig. 6 sweep for the Munich scene.Source: Figure A4 · p. 22
  • Whole-LLM tests report WikiText-2 perplexity and zero-shot accuracy on PIQA, ARC-e, ARC-c, BoolQ, HellaSwag, and WinoGrande.The evaluation runs nine listed LLM variants, with each weight-activation GEMV performed using AIR-LLM’s RF computing.Source: §E.3 · pp. 21–22
  • Airtime is the time one token occupies the wireless link; energy estimates use per-operation costs for converters, digital computing, and data movement.Remote-storage baselines use a standard-compliant 5G NR link, while local-storage baselines occupy no airtime.Source: §E.3 · p. 22

F.1 Benchmarking on GEMV

The Munich sweep shows that AIR-LLM’s energy–accuracy trade-off depends on array size and scene; its energy and airtime per-token plots cover nine LLMs.

F.1 Benchmarking on GEMV

  • For 3-bit ENOB, Munich’s minimum energy costs at A=2/4/8/16/32 are 1.02/2.17/4.32/6.55/6.46×, versus Paris’s 1.02/1.90/3.86/5.09/9.60×.The ordering across array sizes reverses between the two scenes.Source: §F.1 · p. 22
  • In Munich, saturation ENOB rises from 3.13 bits at A=4 to 3.92, 4.71, and 5.24 bits at A=8/16/32; arrays below A=16 cannot reach 4 bits.At A=32, 4-bit ENOB is reached more cheaply than at A=16.Source: §F.1 · p. 22
  • Munich’s 2-bit minimum-energy costs at A=2/4/8/16/32 are 1.02/1.36/2.62/3.85/3.61×.The 4-bit target is out of reach below A=16; at A=16 and 32 it costs 8.52× and 8.08×.Source: §F.1 · p. 22
  • The two scenes show opposite array-size trends because spatial multiplexing needs multipath: Munich gains usable spatial dimensions as the array grows, while Paris runs out.In Paris, saturation ENOB falls from 5.50 bits at A=4 to 4.33–4.62 bits at larger arrays.Source: §F.1 · p. 22
  • Figure A5 plots normalized energy and airtime per token across nine LLMs in Munich.It repeats the energy and airtime comparisons of Figs. 8 and 9.Source: Figure A5 · p. 22

F.2 Benchmarking on LLM

In Munich, AIR-LLM substantially reduces energy relative to storage-based baselines, while its airtime comparison depends on the baseline. LLM outputs remain fluent at higher ENOB and degrade as resolution falls, while zero-shot accuracy varies across tasks.

F.2 Benchmarking on LLM

  • On Munich LLaMA-3.1-8B, AIR-LLM uses 165.5×/32.1× less energy than remote FP16/W3A16 and 51.5×/10.7× less than local baselines.Across nine LLMs, the energy advantage over the two W3A16 baselines remains 28.3–35.3× and 10.0–11.1×.Source: §F.2 · p. 22
  • On Munich, remote FP16 airtime is 5.37–5.42× AIR-LLM’s, while the quantized baseline is only 1.01–1.02× its airtime.The quantized baseline closes from the 1.28× W4A16 ratio in Paris.Source: §F.2 · p. 22
  • Above about 3.5 bits, continuations are fluent; below that, they remain grammatical but drift into invented details and repetition.Sixteen of 20 Paris devices and eight of 20 Munich devices reach above about 3.5 bits.Source: §F.2 · p. 22
  • Paris device 11’s near-singular channel degenerates into repeated tokens at 2.00 bits, producing an outlying PPL of 255.27; failure reflects poor conditioning rather than distance alone.Source: §F.2 · p. 24
  • The worst outcomes track poorly conditioned channels rather than distance alone: Munich’s farthest device still resolves 4.19 bits and has PPL 6.60.The farthest device is 105.0 dB from the central radio.Source: §F.2 · p. 24
  • AIR-LLM averages 68.72% in Paris and 65.67% in Munich on six tasks, versus 75.00% for FP16; the results fall within or near the W3A16 band.AIR-LLM is inside the Paris W3A16 band of 66.00–69.96% and 0.33 points below Munich’s weakest member.Source: §F.2 · p. 24
  • HellaSwag and BoolQ results stay relatively close to FP16, while WinoGrande is lower: 57.77%/57.14% versus 73.56%.The passage attributes WinoGrande’s sensitivity to its narrow candidate log-likelihood margins.Source: §F.2 · p. 24

F.3 Impact of the System Factors

System-factor sweeps show that repetition offers limited returns relative to transmit power, while carrier frequency trades airtime against coverage. In Munich, broadcasting retains a large airtime advantage over unicast at 20 users.

F.3 Impact of the System Factors

  • Sixteen repetitions improve Munich PPL only from 7.37 to 7.19, or 2.4%, while remaining within 0.3% of its 7.17 distortion floor.Repetition lowers thermal noise but costs proportional airtime and cannot remove receiver distortion; transmit power buys 7.9% over the same 10 dB without airtime cost.Source: §F.3 · pp. 24–25, Table A9 · p. 25
  • At 915 MHz, Munich reads 7.28 at 30 dBm versus 7.37 at 50 dBm, but 25 MHz bandwidth makes each GEMV occupy four times the airtime.Both devices remain usable within the regulatory ceiling; Paris has already saturated at 6.72, and the respective floors are 6.69 in Paris and 6.44 in Munich.Source: §F.3 · p. 25, Table A10 · p. 26
  • At 28 GHz, Paris saturates 9% worse at a floor of 7.05 versus 6.48, while Munich remains unusable at the regulatory ceiling.Sparse multipath worsens the per-subcarrier channel conditioning for calibration; the 400 MHz bandwidth cuts airtime fourfold only for devices close enough to the radio.Source: §F.3 · p. 25, Table A10 · p. 26

F.4 Broadcasting to Massive Edge Users

  • At 20 users in Munich, AIR-LLM uses 104.2× and 19.5× less airtime than the FP16 and W3A16 baselines, respectively.Unicast airtime grows linearly with user count while the single broadcast stays flat; the smaller W3A16 gain reflects fewer bits per weight, not a weaker broadcast advantage.Source: §F.4 · p. 26

Related papers

Loading 2610.00465v1…