deep learning toolbox tm  (MathWorks Inc)

Journal: Sensors (Basel, Switzerland)

Article Title: Hardware Implementations of a Deep Learning Approach to Optimal Configuration of Reconfigurable Intelligence Surfaces

doi: 10.3390/s24030899

Figure Lengend Snippet: ROCK 4C Plus, NVIDIA Jetson Nano, Google Coral, and Intel ® Arria ® 10 SX SoC Development Kit specification summary.

Article Snippet: However, as it is not available for any of the other CPU-based platforms or the MATLAB ® Deep Learning HDL Toolbox TM , only implementations from the FP32 model are considered in this case, with the A10_Performance and A10_Generic architectures [ ].

Techniques:

Journal: Sensors (Basel, Switzerland)

Article Title: Hardware Implementations of a Deep Learning Approach to Optimal Configuration of Reconfigurable Intelligence Surfaces

doi: 10.3390/s24030899

Figure Lengend Snippet: NN implementation workflow for the Intel ® Arria ® 10 SX SoC Development Kit device using the MATLAB ® Deep Learning HDL Toolbox TM .

Article Snippet: However, as it is not available for any of the other CPU-based platforms or the MATLAB ® Deep Learning HDL Toolbox TM , only implementations from the FP32 model are considered in this case, with the A10_Performance and A10_Generic architectures [ ].

Techniques:

Journal: Sensors (Basel, Switzerland)

Article Title: Hardware Implementations of a Deep Learning Approach to Optimal Configuration of Reconfigurable Intelligence Surfaces

doi: 10.3390/s24030899

Figure Lengend Snippet: Resource usage in Intel ® Arria ® 10 SX SoC Development Kit with MATLAB ® Deep Learning HDL Toolbox TM .

Article Snippet: However, as it is not available for any of the other CPU-based platforms or the MATLAB ® Deep Learning HDL Toolbox TM , only implementations from the FP32 model are considered in this case, with the A10_Performance and A10_Generic architectures [ ].

Techniques:

Journal: Sensors (Basel, Switzerland)

Article Title: Hardware Implementations of a Deep Learning Approach to Optimal Configuration of Reconfigurable Intelligence Surfaces

doi: 10.3390/s24030899

Figure Lengend Snippet: Inference example, Intel ® Arria ® 10 SX SoC Development Kit with Matlab ® Deep Learning HDL Toolbox TM FP32 accelerator: ( a ) expected RIS, ( b ) inferred RIS, ( c ) error when matching the expected RIS and the inferred RIS, and ( d ) error when matching the opposite of the expected RIS and the inferred RIS (errors are shown in red in both ( c , d ), coincidences in green).

Article Snippet: However, as it is not available for any of the other CPU-based platforms or the MATLAB ® Deep Learning HDL Toolbox TM , only implementations from the FP32 model are considered in this case, with the A10_Performance and A10_Generic architectures [ ].

Techniques:

Journal: Sensors (Basel, Switzerland)

Article Title: Hardware Implementations of a Deep Learning Approach to Optimal Configuration of Reconfigurable Intelligence Surfaces

doi: 10.3390/s24030899

Figure Lengend Snippet: Resource usage in Intel ® Arria ® 10 SX SoC Development Kit with MATLAB ® Deep Learning HDL Toolbox TM .

Article Snippet: Their overall throughput is clearly superior to any of the other alternatives, with an approximately × 20 increase in performance when the Intel ® FPGA AI Suite A10_Performance architecture is compared with the MATLAB ® Deep Learning HDL Toolbox TM FP32 implementation or the NVIDIA Jetson Nano.

Techniques:

Journal: Sensors (Basel, Switzerland)

Article Title: Hardware Implementations of a Deep Learning Approach to Optimal Configuration of Reconfigurable Intelligence Surfaces

doi: 10.3390/s24030899

Figure Lengend Snippet: Inference example, Intel ® Arria ® 10 SX SoC Development Kit with Matlab ® Deep Learning HDL Toolbox TM FP32 accelerator: ( a ) expected RIS, ( b ) inferred RIS, ( c ) error when matching the expected RIS and the inferred RIS, and ( d ) error when matching the opposite of the expected RIS and the inferred RIS (errors are shown in red in both ( c , d ), coincidences in green).

Article Snippet: Their overall throughput is clearly superior to any of the other alternatives, with an approximately × 20 increase in performance when the Intel ® FPGA AI Suite A10_Performance architecture is compared with the MATLAB ® Deep Learning HDL Toolbox TM FP32 implementation or the NVIDIA Jetson Nano.

Techniques:

Journal: Sensors (Basel, Switzerland)

Article Title: Hardware Implementations of a Deep Learning Approach to Optimal Configuration of Reconfigurable Intelligence Surfaces

doi: 10.3390/s24030899

Figure Lengend Snippet: Accuracy comparison of NN execution across the different devices and implementations.

Article Snippet: Their overall throughput is clearly superior to any of the other alternatives, with an approximately × 20 increase in performance when the Intel ® FPGA AI Suite A10_Performance architecture is compared with the MATLAB ® Deep Learning HDL Toolbox TM FP32 implementation or the NVIDIA Jetson Nano.

Techniques: Comparison

Journal: Sensors (Basel, Switzerland)

Article Title: Hardware Implementations of a Deep Learning Approach to Optimal Configuration of Reconfigurable Intelligence Surfaces

doi: 10.3390/s24030899

Figure Lengend Snippet: Performance comparison of NN execution across the different devices and implementations.

Article Snippet: Their overall throughput is clearly superior to any of the other alternatives, with an approximately × 20 increase in performance when the Intel ® FPGA AI Suite A10_Performance architecture is compared with the MATLAB ® Deep Learning HDL Toolbox TM FP32 implementation or the NVIDIA Jetson Nano.

Techniques: Comparison

Journal: Sensors (Basel, Switzerland)

Article Title: Hardware Implementations of a Deep Learning Approach to Optimal Configuration of Reconfigurable Intelligence Surfaces

doi: 10.3390/s24030899

Figure Lengend Snippet: Graphical performance comparison for ( a ) FP32 implementations and ( b ) INT8-quantized implementations (performance of the A10_Performance implementation for the Intel ® Arria ® 10 SoC DevKit and Intel ® FPGA AI Suite is shown in green as a benchmark).

Article Snippet: Their overall throughput is clearly superior to any of the other alternatives, with an approximately × 20 increase in performance when the Intel ® FPGA AI Suite A10_Performance architecture is compared with the MATLAB ® Deep Learning HDL Toolbox TM FP32 implementation or the NVIDIA Jetson Nano.

Techniques: Comparison

Journal: Sensors (Basel, Switzerland)

Article Title: Hardware Implementations of a Deep Learning Approach to Optimal Configuration of Reconfigurable Intelligence Surfaces

doi: 10.3390/s24030899

Figure Lengend Snippet: Resource usage comparison in Intel ® Arria ® 10 SX SoC Development Kit.

Article Snippet: Their overall throughput is clearly superior to any of the other alternatives, with an approximately × 20 increase in performance when the Intel ® FPGA AI Suite A10_Performance architecture is compared with the MATLAB ® Deep Learning HDL Toolbox TM FP32 implementation or the NVIDIA Jetson Nano.

Techniques: Comparison

Journal: Sensors (Basel, Switzerland)

Article Title: Hardware Implementations of a Deep Learning Approach to Optimal Configuration of Reconfigurable Intelligence Surfaces

doi: 10.3390/s24030899

Figure Lengend Snippet: ROCK 4C Plus, NVIDIA Jetson Nano, Google Coral, and Intel ® Arria ® 10 SX SoC Development Kit specification summary.

Article Snippet: It is also worth noting the difference in device occupation between the two MATLAB ® Deep Learning HDL Toolbox TM implementations: while the INT8-quantized version implies a slight reduction in accuracy, it is able to almost double the performance over the FP32 option thanks to a more intensive use of ALMs and, particularly, the embedded-multiplier variable-precision DSP blocks.

Techniques:

Journal: Sensors (Basel, Switzerland)

Article Title: Hardware Implementations of a Deep Learning Approach to Optimal Configuration of Reconfigurable Intelligence Surfaces

doi: 10.3390/s24030899

Figure Lengend Snippet: NN implementation workflow for the Intel ® Arria ® 10 SX SoC Development Kit device using the MATLAB ® Deep Learning HDL Toolbox TM .

Article Snippet: It is also worth noting the difference in device occupation between the two MATLAB ® Deep Learning HDL Toolbox TM implementations: while the INT8-quantized version implies a slight reduction in accuracy, it is able to almost double the performance over the FP32 option thanks to a more intensive use of ALMs and, particularly, the embedded-multiplier variable-precision DSP blocks.

Techniques:

Journal: Sensors (Basel, Switzerland)

Article Title: Hardware Implementations of a Deep Learning Approach to Optimal Configuration of Reconfigurable Intelligence Surfaces

doi: 10.3390/s24030899

Figure Lengend Snippet: Resource usage in Intel ® Arria ® 10 SX SoC Development Kit with MATLAB ® Deep Learning HDL Toolbox TM .

Article Snippet: It is also worth noting the difference in device occupation between the two MATLAB ® Deep Learning HDL Toolbox TM implementations: while the INT8-quantized version implies a slight reduction in accuracy, it is able to almost double the performance over the FP32 option thanks to a more intensive use of ALMs and, particularly, the embedded-multiplier variable-precision DSP blocks.

Techniques:

Journal: Sensors (Basel, Switzerland)

Article Title: Hardware Implementations of a Deep Learning Approach to Optimal Configuration of Reconfigurable Intelligence Surfaces

doi: 10.3390/s24030899

Figure Lengend Snippet: Inference example, Intel ® Arria ® 10 SX SoC Development Kit with Matlab ® Deep Learning HDL Toolbox TM FP32 accelerator: ( a ) expected RIS, ( b ) inferred RIS, ( c ) error when matching the expected RIS and the inferred RIS, and ( d ) error when matching the opposite of the expected RIS and the inferred RIS (errors are shown in red in both ( c , d ), coincidences in green).

Article Snippet: It is also worth noting the difference in device occupation between the two MATLAB ® Deep Learning HDL Toolbox TM implementations: while the INT8-quantized version implies a slight reduction in accuracy, it is able to almost double the performance over the FP32 option thanks to a more intensive use of ALMs and, particularly, the embedded-multiplier variable-precision DSP blocks.

Techniques: