HPE machine learning projects

Size: px

Start display at page:

Download "HPE machine learning projects"

Ashlie Sullivan
5 years ago
Views:

1 HPE machine learning projects Chris Cheng, Distinguished Technologist, Yongjin Choi, Master Technologist, HP Enterprise storage division 10/28/2017 1

2 Multi-physics non-linear channel analysis Key objectives Direct generation of non-linear model through machine learning Multi-physics models with neural networks Embedding neural network models into test instruments Cloud centric analysis with massively parallel scalability with neural networks 10/28/2017 2

3 Single step non-linear state space model NNSS = Nonlinear State-Space based on Neural Network EQ3 EQ2 EQ1 EQ4 20 Target System Magnitude (db) Frequency (khz) 3 10/28/2017 3

4 External measurement Silicon Tx Channel Silicon Rx Channel can be copper, optical or multi-physics Supervised system Control stimulus and observe output RNN training 10/28/2017 4

5 Mixed external and internal silicon measureement Adjusting CTLE gain, pk freq, DC gain, offset etc Silicon Tx output Channel Silicon Rx with internal waveform measurement Output from internal measurement Input at silicon pin RNN training 10/28/2017 5

6 Comparison with AMI model 10/28/2017 6

From: u1 To: y1 EQ1 8 EQ2 EQ3 EQ4 6 S-Param 4 Magnitude

7 Application : neuron embedded oscilloscope Raw measured input n(k) EQ4 EQ3 EQ2 EQ1 C(q) x(k) B(q) 1/A(q) y(k) 10 From: u1 To: y1 EQ1 8 EQ2 EQ3 EQ4 6 S-Param 4 Magnitude (db) Frequency (GHz) 10/28/2017 7

8 Deep BER multi-physics non-linear coded channel analysis Deep BER : BER10e-12 to 10e-15 Multi-physics : Mixed of copper, active optical and/or active redriver interconnect Noisy : crosstalk, clock/cdr phase, optical, thermal noise Coded : Reed Solomon FEC code gain to bring raw correctable error down to target uncorrectable BER 10/28/2017 8

Channel Crosstalk Channel Crosstalk Channel parametric sweep Raw error rate Channel loss 0% 10% 12% 13% 14% 15% 16% 18% 0% 1.0E-15 1.0E-15 1.0E-15 1.0E-12 9.1E-09 1.5E-07 7.9E-07 9.0E-06 10% 1.

9 Channel Crosstalk Channel Crosstalk Channel parametric sweep Raw error rate Channel loss 0% 10% 12% 13% 14% 15% 16% 18% 0% 1.0E E E E E E E E-06 10% 1.0E E E E E E E E-05 20% 1.0E E E E E E E E-05 30% 1.0E E E E E E E E-05 40% 1.0E E E E E E E E-05 50% 1.0E E E E E E E E-05 60% 1.0E E E E E E E E-05 70% 1.0E E E E E E E E-05 80% 1.0E E E E E E E E-05 90% 1.0E E E E E E E E % 2.6E E E E E E E E-04 Uncorrectable error rate after FEC Channel loss 0% 10% 12% 13% 14% 15% 16% 18% 0% 1.0E E E E E E E E-08 10% 1.0E E E E E E E E-08 20% 1.0E E E E E E E E-07 30% 1.0E E E E E E E E-07 40% 1.0E E E E E E E E-07 50% 1.0E E E E E E E E-07 60% 1.0E E E E E E E E-07 70% 1.0E E E E E E E E-07 80% 1.0E E E E E E E E-07 90% 1.0E E E E E E E E % 1.0E E E E E E E E-06 BER14/15 BER12/13 BER10/11 BER8/9 BER6/7 BER4/5 0% 10% 12% 13% 14% 15% Coding Gain 16% 18% 100% 80% 60% 40% 20% 0% 1.00E E E E E E E E E E E E+07 10/28/2017 9

10 Cloud based channel analysis : Parallel convergent sweep Tx control neural net Channel condition control neural net EC2 blocks SI Tx SI channel SI Rx SI Tx SI channel SI Rx Rx control neural net Sweep through the Tx/Rx with a fixed topology and converge to optimal eye opening SI Tx SI channel SI Rx Sweeping blocks 10/28/

11 Cloud based channel analysis : Parallel deep PRBS run Break long PRBS patterns during long disparity into smaller simulation sequence SI Tx SI channel SI Rx SI Tx SI channel SI Rx SI Tx SI channel SI Rx 10/28/

12 Cloud based analysis : Putting it together Channel Crosstalk Raw error rate Channel loss 0% 10% 12% 13% 14% 15% 16% 18% 0% 1.0E E E E E E E E-06 10% 1.0E E E E E E E E-05 20% 1.0E E E E E E E E-05 30% 1.0E E E E E E E E-05 40% 1.0E E E E E E E E-05 50% 1.0E E E E E E E E-05 60% 1.0E E E E E E E E-05 70% 1.0E E E E E E E E-05 80% 1.0E E E E E E E E-05 90% 1.0E E E E E E E E % 2.6E E E E E E E E-04 Uncorrectable error rate after FEC Channel loss Channel Condition Tx control neural net Sweep channel condition control neural net EC2 blocks SI Tx SI channel SI Rx SI Tx SI channel SI Rx Rx control neural net Channel Crosstalk 0% 10% 12% 13% 14% 15% 16% 18% 0% 1.0E E E E E E E E-08 10% 1.0E E E E E E E E-08 20% 1.0E E E E E E E E-07 30% 1.0E E E E E E E E-07 40% 1.0E E E E E E E E-07 50% 1.0E E E E E E E E-07 60% 1.0E E E E E E E E-07 70% 1.0E E E E E E E E-07 80% 1.0E E E E E E E E-07 90% 1.0E E E E E E E E % 1.0E E E E E E E E-06 FEC coding correction SI Tx SI channel SI Rx 10/28/

13 Cloud based analysis : Putting it together (cont.) Channel Crosstalk Raw error rate Channel loss 0% 10% 12% 13% 14% 15% 16% 18% 0% 1.0E E E E E E E E-06 10% 1.0E E E E E E E E-05 20% 1.0E E E E E E E E-05 30% 1.0E E E E E E E E-05 40% 1.0E E E E E E E E-05 50% 1.0E E E E E E E E-05 60% 1.0E E E E E E E E-05 70% 1.0E E E E E E E E-05 80% 1.0E E E E E E E E-05 90% 1.0E E E E E E E E % 2.6E E E E E E E E SI Tx SI channel SI Rx SI Tx SI channel SI Rx SI Tx SI channel SI Rx Channel Crosstalk Uncorrectable error rate after FEC Channel loss 0% 10% 12% 13% 14% 15% 16% 18% 0% 1.0E E E E E E E E-08 10% 1.0E E E E E E E E-08 20% 1.0E E E E E E E E-07 30% 1.0E E E E E E E E-07 40% 1.0E E E E E E E E-07 50% 1.0E E E E E E E E-07 60% 1.0E E E E E E E E-07 70% 1.0E E E E E E E E-07 80% 1.0E E E E E E E E-07 90% 1.0E E E E E E E E % 1.0E E E E E E E E For every channel condition point 1. Sweep the equalization models to find the convergence 2. Run deep PRBS pattern in parallel to simulate raw error pattern 3. Perform FEC to find uncorrectable errors 10/28/

14 Self correcting models, tie in with measurement Component vendor provide the original neural net model Test instrument perform measurements If the measurements have significant deviations, retrain neural network to match model Advantages Complete IP protection Self correcting Better accuracy than IBIS AMI Hardware acceleration Lots of free tool kits available Components vendor SI tool Test instruments Reference SI model (initial RNN) Observed SI model (Retrained RNN) CAD supplier Designer 10/28/

15 System Analytics : Proactive failure prediction Failure prediction based on industrial standard signals SMART (Self-Monitoring, Analysis and Reporting Technology) bus ID # Attribute Name 5 Reallocated Sectors Count 187 Reported Uncorrectable Errors 188 Command Timeout 197 Current pending Sector Count 198 Offline Uncorrectable SMBus ASF sensor readings Voltage Fan speed Temperature Current 10/28/2017 HPE 15

16 Using ANN for classification Feedforward multi-layer perceptron Non-parametric feature selection SMART ID# 5 SMART ID#187 SMART ID #** Input layer Good/Bad Output layer Hidden layer 10/28/

17 Receiver operating characteristics (ROC) plots Failure Prediction with ANN (%) 10/28/

18 Failure lead time Window N =1 : Window N =3 : Window N =5 : Average Lead Time =14.1 days Average Lead Time =7.3 days Average Lead Time =6.28 days 10/28/

19 Failure prediction model t d t w t l t p t t = present time t d = observation window (sample history require to generate prediction) t w = alert time (time require to receive, process and generate prediction) t l = lead time (time to failure) t r = response time t p = error range (range of failure before system correction) t r time Beyond hardware failure Software/system failures Syslog/Sys Activity report 10/28/

TITLE. Novel Methodology of IBIS-AMI Hardware Correlation using Trend and Distribution Analysis for high-speed SerDes System

TITLE. Novel Methodology of IBIS-AMI Hardware Correlation using Trend and Distribution Analysis for high-speed SerDes System TITLE Novel Methodology of IBIS-AMI Hardware Correlation using Trend and Distribution Analysis for high-speed SerDes System Hong Ahn, (Xilinx) Brian Baek, (Cisco) Ivan Madrigal (Xilinx) Image Hongtao Zhang