![Page 1: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/1.jpg)
1
Ahmed Alkhateeb, Assistant Professor, Arizona State University
Adriana Flores Miranda, NVIDIA
Nima PourNejatian, NVIDIA
5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS
![Page 2: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/2.jpg)
2
Introduction
5G key aspects
MIMO limitations
Why Deep Learning overcomes limitations
Deep Learning in 5G applications
Auto-Precoder
Environment-aware joint channel estimation and precoding for mmWave MIMO
Demo and Results
An example based on accurate 3D ray-tracing simulations
AGENDA
![Page 3: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/3.jpg)
3
TREND: 4G IS MATURING WHILE 5G IS UNDER WAY
296 operators in 100 countries that have been investing in 5G
Global Wireless Telecommunications Carriers Industry
39 operators with 5G launches (either mobile or FWA, some with limited availability)
Source: GSA, August 2019
![Page 4: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/4.jpg)
4
7M+ 1M 100 Gbps < 1msMacro 4G Base Stations to be upgraded to 5G
IoT Devices / KM2 Bandwidth Latency= AI at the Edge
5G OPPORTUNITY
![Page 5: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/5.jpg)
5
TELCO’S CHALLENGES5G mmWave, Massive MIMO, and AI
STRATEGIESCHALLENGE
✓ 5G, mmWave, and Massive MIMO
✓ Artificial Intelligence
✓ Software Defined Networks
✓ Edge Computing
Traffic per Customer
Cost per Customer
Revenue per Customer
![Page 6: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/6.jpg)
6
WHY MASSIVE MIMO AND 5G MMWAVEBenefits And Requirement
5G mmWave, Massive MIMO signal
processing while meeting the low
latency requirement
Higher FLOPS Lower FLOPS
Lower Latency
MIM
O L
ayer
User data rate per channel (bps) is limited by:
𝑅 ≤ 𝑁𝑀𝐼𝑀𝑂 𝐵𝑊𝑁𝑢𝑠𝑒𝑟𝑠
𝑙𝑜𝑔2(1+𝑆𝑁𝐼𝑅)
▪ Massive MIMO: Higher 𝑁𝑀𝐼𝑀𝑂▪ 5G mmWave:
▪ Higher 𝐵𝑊 (from 20 MHz in 4G to 800 MHz in 5G
mmWave )
▪ Directional Beamforming:
▪ Higher 𝑆𝑁𝐼𝑅▪ Lower 𝑁𝑢𝑠𝑒𝑟𝑠 sharing the beam resources
How 5G enables 10Gbps+ data rates?
Key Objective
![Page 7: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/7.jpg)
7
CHALLENGES WITH SCALING UP MIMO IN 5G AND BEYONDChannel acquisition and hardware power consumption
![Page 8: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/8.jpg)
8
DEEP LEARNING IN 5GDL can overcome MIMO limitations
We propose to leverage ML models to learn this mapping function
![Page 9: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/9.jpg)
9
DEEP LEARNING APPLICATIONS IN 5GDeep learning enables reliable and highly-mobile massive MIMO applications
![Page 10: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/10.jpg)
10
DEEP LEARNING APPLICATIONS IN 5GDeep learning enables reliable and highly-mobile mmWave applications
![Page 11: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/11.jpg)
11
DEEP LEARNING APPLICATIONS IN 5GDeep learning supports practical large-scale MIMO transceivers
![Page 12: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/12.jpg)
12
AUTO-PRECODER
![Page 13: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/13.jpg)
13
BACKGROUND AND MOTIVATIONWhy hybrid analog-digital architectures?
Hybrid analog/digital architectures achieve high data rates with reasonable complexity
+
wRF
+
wRFww
RF
Combiner
N
RF Chain + ADC
rx
Analog-only Hybrid analog/digitalFully-digital
![Page 14: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/14.jpg)
14
BACKGROUND AND MOTIVATIONChannel estimation is challenging!
Channel is seen through the RF lens
Analog circuits add strict constraints
Leveraging hybrid architectures requires developing efficient channel estimation solutions
![Page 15: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/15.jpg)
15
BACKGROUND AND MOTIVATIONClassical channel estimation approaches for hybrid architectures
Sensing the channel with random beam patterns
Sparse channel reconstruction using approaches such as OMP
Random beams sense directions that may never be used
Prior channel observations are not leveraged
Classical Compressive Sensing Approach
Limitations
![Page 16: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/16.jpg)
16
BACKGROUND AND MOTIVATIONClassical channel estimation approaches for hybrid architectures
Sensing the channel with random beam patterns
Sparse channel reconstruction using approaches such as OMP
Random beams sense directions that may never be used
Prior channel observations are not leveraged
Classical Compressive Sensing Approach
LimitationsHow can deep learning help?
![Page 17: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/17.jpg)
17
AUTO PRECODERKey Idea
mmWave channel estimation .. followed by hybrid precoding design
Neural network weights realize measurement beams focusing on important directions
Channel reconstruction leverages prior observation
Single-layer
Perceptron
“Channel Estimator” Neural Network
Sparse Reconstruction Algorithm Ex: OMP
BeamformingDesign
![Page 18: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/18.jpg)
18
AUTO PRECODERKey Idea
BeamformingDesign
Sparse Reconstruction Algorithm Ex: OMP
Single-layer
Perceptron
mmWave channel estimation .. followed by hybrid precoding design
Proposed “Auto-precoder”: Optimizes measurements and leverages prior observations
“Precoder” Neural Network
![Page 19: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/19.jpg)
19
AUTO PRECODERJoint channel sensing and precoder prediction
BeamformingDesign
Sparse Reconstruction Algorithm Ex: OMP
Channel sensingwith & .
![Page 20: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/20.jpg)
20
AUTO PRECODERJoint channel sensing and precoder prediction
BeamformingDesign
Sparse Reconstruction Algorithm Ex: OMP
Channel sensingwith & .
“channel sensing” Neural Network
“Beam Prediction” Neural Network
![Page 21: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/21.jpg)
21
HYBRID BEAM PREDICTIONJoint channel sensing and hybrid beam prediction
Learns optimized
measurements at
both TX and RX
Predicts TX and RX
hybrid beams
X. Li, and A. Alkhateeb "Deep Learning for Direct Hybrid Precoding in Millimeter Wave Massive MIMO Systems” Asilomar 2019 (arXiv: https://arxiv.org/abs/1905.13212)
![Page 22: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/22.jpg)
22
REAL WORLD DEPLOYMENT
![Page 23: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/23.jpg)
23
THE NEED FOR RAY-TRACING
Studying the performance of the proposed deep learning approaches needs channel datasets
Generated channels should capture the dependency on the environment
Accurate 3D ray-tracing simulators could be the solution
![Page 24: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/24.jpg)
24
REMCOM WIRELESS INSITE: AN ACCURATE RAY-TRACING TOOL
Accurate 3D ray-tracingAdvanced propagation
modelsMIMO capabilities
Ray-tracing results have been validated with measurements at both sub-6GHz and
mmWave
![Page 25: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/25.jpg)
25
DEEPMIMO: A DEEP-LEARNING DATASET FOR MIMO SYSTEMShttps://www.deepmimo.net/
The DeepMIMO dataset enables a wide range of machine learning tasks
![Page 26: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/26.jpg)
26
1-DATA COLLECTION: MMWAVE PRECODING PREDICTIONData Collection and Evaluation Using Accurate 3D Ray Tracing
![Page 27: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/27.jpg)
27
1-DATA COLLECTION: MMWAVE PRECODING PREDICTIONData Collection and Evaluation Using Accurate 3D Ray Tracing
Top View
mmWave BS
mmWave BS
![Page 28: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/28.jpg)
28
1-DATA COLLECTION: MMWAVE PRECODING PREDICTIONData Collection and Evaluation Using Accurate 3D Ray Tracing
Top View
mmWave BS
![Page 29: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/29.jpg)
29
1-DATA COLLECTION: MMWAVE PRECODING PREDICTIONData Collection and Evaluation Using Accurate 3D Ray Tracing
Top View
mmWave BS
![Page 30: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/30.jpg)
30
1-DATA COLLECTION: MMWAVE PRECODING PREDICTIONData Collection and Evaluation Using Accurate 3D Ray Tracing
Top View
mmWave BS
![Page 31: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/31.jpg)
31
1-DATA COLLECTION: MMWAVE PRECODING PREDICTIONData Collection and Evaluation Using Accurate 3D Ray Tracing
Top View
mmWave BS
![Page 32: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/32.jpg)
32
1-DATA COLLECTION: MMWAVE PRECODING PREDICTIONData Collection and Evaluation Using Accurate 3D Ray Tracing
Top View
mmWave BS
![Page 33: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/33.jpg)
33
1-DATA COLLECTION: MMWAVE PRECODING PREDICTIONData Collection and Evaluation Using Accurate 3D Ray Tracing
Top View
Estimated Channel BS Beams Mobile User Beams
Hybrid PrecodingDesign
Dataset Construction
mmWave BS
![Page 34: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/34.jpg)
34
1-DATA COLLECTION: MMWAVE PRECODING PREDICTIONData Collection and Evaluation Using Accurate 3D Ray Tracing
Top View
Estimated Channel BS Beams Mobile User Beams
Hybrid PrecodingDesign
Dataset Construction
mmWave BS
![Page 35: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/35.jpg)
35
1-DATA COLLECTION: MMWAVE PRECODING PREDICTIONData Collection and Evaluation Using Accurate 3D Ray Tracing
Top View
Estimated Channel BS Beams Mobile User Beams
Hybrid PrecodingDesign
Dataset Construction
mmWave BS
![Page 36: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/36.jpg)
36
1-DATA COLLECTION: MMWAVE PRECODING PREDICTIONData Collection and Evaluation Using Accurate 3D Ray Tracing
Top View
Estimated Channel BS Beams Mobile User Beams
Hybrid PrecodingDesign
Dataset Construction
mmWave BS
![Page 37: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/37.jpg)
37
1-DATA COLLECTION: MMWAVE PRECODING PREDICTIONData Collection and Evaluation Using Accurate 3D Ray Tracing
Top View
Estimated Channel BS Beams Mobile User Beams
Hybrid PrecodingDesign
Dataset Construction
mmWave BS
![Page 38: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/38.jpg)
38
2. TRAINING: MMWAVE PRECODING PREDICTIONTraining the Auto-Precoder Neural Network Model
Top View
mmWave BS
The collected dataset is used to train the NN end-to-end
![Page 39: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/39.jpg)
39
2. TRAINING: MMWAVE PRECODING PREDICTIONTraining the Auto-Precoder Neural Network Model
Top View
mmWave BS
The collected dataset is used to train the NN end-to-end
![Page 40: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/40.jpg)
40
2. TRAINING: MMWAVE PRECODING PREDICTIONTraining the Auto-Precoder Neural Network Model
Top View
mmWave BS
The collected dataset is used to train the NN end-to-end
![Page 41: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/41.jpg)
41
2. TRAINING: MMWAVE PRECODING PREDICTIONTraining the Auto-Precoder Neural Network Model
Top View
mmWave BS
The collected dataset is used to train the NN end-to-end
![Page 42: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/42.jpg)
42
2. TRAINING: MMWAVE PRECODING PREDICTIONTraining the Auto-Precoder Neural Network Model
Top View
mmWave BS
The collected dataset is used to train the NN end-to-end
![Page 43: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/43.jpg)
43
2. TRAINING: MMWAVE PRECODING PREDICTIONTraining the Auto-Precoder Neural Network Model
Top View
mmWave BS
The collected dataset is used to train the NN end-to-end
![Page 44: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/44.jpg)
44
2. TRAINING: MMWAVE PRECODING PREDICTIONTraining the Auto-Precoder Neural Network Model
Top View
mmWave BS
The collected dataset is used to train the NN end-to-end
![Page 45: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/45.jpg)
45
RESULTS
![Page 46: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/46.jpg)
46
TRANSMIT BEAM ACCURACY90%+ Accuracy in Beam Perdition with A Few Pilots
48
%
59
%
60
% 66
%
59
% 68
%
68
%
79
%
65
%
76
% 80
% 84
%
74
%
83
% 87
%
87
%
79
% 87
%
89
%
90
%
81
%
90
%
91
%
92
%
84
% 90
%
93
%
93
%
4 9 16 64
TRA
NSM
IT A
CC
UR
AC
Y
NO. PILOTS
Pt = 0 Pt = 5 Pt = 10 Pt = 15 Pt = 20 Pt = 25 Pt = 30 dBm
![Page 47: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/47.jpg)
47
RECEIVER BEAM ACCURACY90%+ Accuracy in Beam Perdition with A Few Pilots
55
% 60
%
62
%
65
%
63
% 68
%
71
%
82
%
69
%
78
% 81
% 86
%
77
% 86
%
88
%
89
%
81
% 89
%
91
%
92
%
83
% 92
%
93
%
93
%
86
% 92
%
94
%
94
%
4 9 16 64
REC
EIV
ER A
CC
UR
AC
Y
NO. PILOTS
Pt = 0 Pt = 5 Pt = 10 Pt = 15 Pt = 20 Pt = 25 Pt = 30 dBm
![Page 48: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/48.jpg)
48
ACHIEVABLE DATA RATESPerformance: With a few measurements, 4X higher data rate
4X
![Page 49: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/49.jpg)
49
TRAINING TIMEParameters: Batch Size of 64 and 15 Epoch Counts
2,493
119
CPU: Xeon E5-2630 v4 @2.2GHz GPU: Tesla V100-DGXS-16GB
Training Time (seconds)
20X
Speedup!
![Page 50: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/50.jpg)
50
INFERENCE TIMEPer each UE
11.77
1.50
0.31
TRADITIONAL(CHANNEL ESTIMATION,
PRECODING FOR 8 STREAMS)
DEEP LEARNING: AUTO-PRECODERCPU (XEON E5-2630 V4 @2.2GHZ)
DEEP LEARNING: AUTO-PRECODERGPU (TESLA V100-DGXS-16GB)
Inference Time (msec)
![Page 51: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/51.jpg)
51
TAKEAWAYS
• 5G mmwave massive MIMO is promising but have limitations
• Channel acquisition overhead is a key challenge
• Deep Learning has the potential of predicting these channel and beams and thus removing the overhead
• Ray tracing is needed to construct realistic and accurate dataset
• Our proposed solution – prior channel observations to optimize the sensing beams to focus where the users are and predict the beams effectively without channel estimation
• The proposed hybrid beam prediction outperforms traditional methods with few measurements
• This is done in less time than traditional methods thanks to accelerated inference on GPUs.
5G Meets Deep Learning, Ray Tracing, and GPUs
![Page 52: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/52.jpg)
52
QUESTIONS
•Ahmed Alkhateeb: [email protected]
•Adriana Flores: [email protected]
•Nima Pour Nejatian: [email protected]
How to contact us
![Page 53: 5G MEETS DEEP LEARNING, RAY TRACING, AND GPUS...Why Deep Learning overcomes limitations Deep Learning in 5G applications Auto-Precoder Environment-aware joint channel estimation and](https://reader036.vdocuments.net/reader036/viewer/2022062609/60f7e43901610563f30b2ed7/html5/thumbnails/53.jpg)