Skip to content

ip_core_comparison.md

Latest commit

 

History

History
323 lines (236 loc) · 9.22 KB

ip_core_comparison.md

File metadata and controls

323 lines (236 loc) · 9.22 KB

IP Core Comparison Tool

Date: 2025-10-24 Task: Create comprehensive IP core comparison tool for SoC integration Tool: cli/compare_ip_cores.py

Summary

Created a new CLI tool that compares licensable AI/compute IP cores for custom SoC integration, enabling hardware architects to evaluate IP licensing options before tape-out.

New IP Core: ARM Mali-G78 MP20

Added ARM Mali-G78 MP20 GPU IP mapper:

Specifications:

  • Performance: 1.94 TFLOPS FP32 @ 848 MHz (~97 GFLOPS per core)
  • Architecture: 20 shader cores, 2nd gen Valhall, unified shader (graphics+compute)
  • Power: 5W typical TDP
  • Precision: FP32 (native), FP16 (2× FP32), INT8 (not optimized)
  • Use Cases: Mobile gaming, computational photography, light AI inference
  • Real-world: Used in Google Tensor (Pixel 6/6 Pro)

Factory Function: create_arm_mali_g78_mp20_mapper() Resource Model: arm_mali_g78_mp20_resource_model()

IP Cores Comparison Tool

Purpose

Compare 6 licensable IP cores across 2 categories for SoC integration decisions:

Category 1: Neural Processing IP (≤5W)

  • CEVA NeuPro-M NPM11: 20 TOPS INT8 @ 2W
  • Cadence Tensilica Vision Q8: 3.8 TOPS INT8 @ 1W
  • Synopsys ARC EV7x: 35 TOPS INT8 @ 5W
  • ARM Mali-G78 MP20: 1.94 TFLOPS FP32 @ 5W

Category 2: High-Performance IP (≤30W)

  • KPU-T64: 6.9 TOPS INT8 @ 6W
  • KPU-T256: 33.8 TOPS INT8 @ 30W

Key Features

  1. SoC Integration Focus

    • Process node recommendations (5nm/7nm/16nm)
    • Typical use case mapping
    • Licensing considerations
    • Integration timeline estimates

  2. Performance Metrics

    • Peak TOPS/TFLOPS
    • Latency and throughput (FPS)
    • Power efficiency (FPS/W, TOPS/W)
    • Energy per inference
    • Utilization analysis

  3. Test Models

    • ResNet-50: Vision backbone
    • DeepLabV3+: Semantic segmentation
    • ViT-Base: Vision Transformer

Benchmark Results (ResNet-50 @ INT8)

Category 1: Neural Processing IP

IP Core Vendor Power Latency FPS FPS/W Utilization
CEVA NeuPro-M NPM11 CEVA 2W 150.6 ms 6.6 3.32 29.3%
Cadence Vision Q8 Cadence 1W 225.3 ms 4.4 4.44 ⭐ 47.7%
Synopsys ARC EV7x Synopsys 5W 364.1 ms 2.7 0.55 14.7%
ARM Mali-G78 MP20 ARM 5W 1221.8 ms 0.8 0.16 99.2%

Key Insights:

  • Best FPS/W: Cadence Vision Q8 (4.44 FPS/W) - Lowest power, vision-optimized
  • Best Latency: CEVA NeuPro-M (150.6 ms) - Highest TOPS/W (10 TOPS/W)
  • Highest Peak TOPS: Synopsys ARC EV7x (35 TOPS) - Automotive-grade
  • Graphics GPU: ARM Mali-G78 slower for AI workloads (graphics-optimized)

Category 2: High-Performance IP

IP Core Vendor Power Latency FPS FPS/W Utilization
KPU-T64 KPU 6W 4.2 ms 238.8 39.79 ⭐ 98.8%
KPU-T256 KPU 30W 1.1 ms ⭐ 893.2 29.77 90.9%

Key Insights:

  • Best FPS/W: KPU-T64 (39.79 FPS/W) - Best efficiency
  • Best Latency: KPU-T256 (1.1 ms) - Highest throughput

SoC Integration Recommendations

Use Case Mapping

Use Case Recommended IP Process Node Rationale
Mobile Flagship CEVA NeuPro @ 7nm 5nm/7nm Highest TOPS/W, mobile-optimized
Mobile Camera ISP Cadence Vision Q8 7nm Vision pipeline integration, 1W
Automotive ADAS Synopsys ARC EV7x 16nm/7nm Automotive-grade, 35 TOPS
Mobile Gaming ARM Mali-G78 MP20 7nm/5nm Graphics+compute hybrid
Edge Server KPU-T256 16nm Highest performance, 33.8 TOPS

Process Node Guidance

  • 5nm: Latest mobile SoCs (Apple A-series, Snapdragon 8 Gen)

    • Best: CEVA NeuPro, ARM Mali-G78
    • Power: Lowest, highest cost

  • 7nm: Mainstream mobile/automotive (Google Tensor, Exynos)

    • Best: CEVA NeuPro, Cadence Vision Q8, Synopsys ARC EV7x
    • Power: Good balance, moderate cost

  • 16nm: Cost-effective automotive/IoT

    • Best: Synopsys ARC EV7x, KPU-T64/T256
    • Power: Higher, lowest cost

Licensing Considerations

  1. Upfront Licensing Fees

    • One-time fee per IP core license
    • Varies by vendor ($100K - $10M+)
    • Volume discounts available

  2. Royalty Models

    • Per-chip royalty (% of chip price)
    • Per-unit royalty (fixed per chip)
    • Royalty caps and thresholds

  3. Integration Timeline

    • 6-12 months: SoC integration and tape-out
    • 3-6 months: Verification and validation
    • Additional: Automotive certification (if required)

  4. Support and Ecosystem

    • Compiler and toolchain
    • Reference designs
    • Technical support
    • Training and documentation

Files Modified/Created

New Files

  1. src/graphs/characterize/hardware_mapper.py

    • Added arm_mali_g78_mp20_resource_model() (lines 4497-4668)
    • +172 lines

  2. src/graphs/characterize/gpu_mapper.py

    • Added create_arm_mali_g78_mp20_mapper() (lines 359-414)
    • +56 lines

  3. cli/compare_ip_cores.py (new file)

    • Comprehensive IP core comparison tool
    • +570 lines
    • Categories: Neural Processing IP (≤5W), High-Performance IP (≤30W)
    • Executive summary with SoC integration guidance

  4. docs/IP_CORE_COMPARISON_TOOL.md (this file)

    • Documentation of new tool and ARM Mali mapper

Modified Files

  1. cli/README.md
    • Added documentation for compare_ip_cores.py
    • Usage examples and output samples
    • SoC integration guidance summary

Usage

Basic Usage

# Run full IP core comparison
python cli/compare_ip_cores.py

Output Structure

  1. Category 1 Results: Neural Processing IP (≤5W)

    • ResNet-50, DeepLabV3+, ViT-Base benchmarks
    • Performance table with latency, FPS, FPS/W

  2. Category 2 Results: High-Performance IP (≤30W)

    • Same models, higher performance IP cores
    • KPU-T64 and KPU-T256 comparison

  3. Executive Summary

    • Best by metric analysis
    • Vendor comparison
    • SoC integration recommendations
    • Licensing considerations

Classification: IP Cores vs. Platforms

IP Cores (Licensable)

✅ CEVA NeuPro-M NPM11 ✅ Cadence Tensilica Vision Q8 ✅ Synopsys ARC EV7x ✅ ARM Mali-G78 MP20 ✅ KPU-T64 (can be licensed) ✅ KPU-T256 (can be licensed)

Characteristics:

  • Licensable for SoC integration
  • Configurable (cores, frequency, cache)
  • Requires customer tape-out
  • Flexible process node
  • Customer owns the silicon

Complete Platforms (Not IP Cores)

❌ TI TDA4VM - Complete automotive SoC ❌ Qualcomm QRB5165 - Complete robotics platform ❌ NVIDIA Jetson - Complete edge AI platform ❌ Hailo-8/10H - Complete AI accelerator chip

Characteristics:

  • Fixed silicon product
  • Vendor manufactures and sells
  • Fixed configuration
  • Pre-integrated ecosystem
  • Customer buys the chip

Performance Analysis

Best Overall (ResNet-50 @ INT8)

Power Efficiency: Cadence Vision Q8

  • 4.44 FPS/W
  • Only 1W power
  • Best for always-on applications

Absolute Performance: KPU-T256

  • 1.1 ms latency
  • 893 FPS throughput
  • Best for edge servers

Balanced: CEVA NeuPro-M NPM11

  • 10 TOPS/W efficiency
  • 150 ms latency
  • Best for mobile AI

Automotive: Synopsys ARC EV7x

  • 35 TOPS peak
  • Automotive-grade architecture
  • Best for ADAS

Graphics+AI: ARM Mali-G78 MP20

  • 1.94 TFLOPS FP32
  • Mobile gaming optimized
  • Use with dedicated NPU for AI

Future Work

  1. Additional IP Cores

    • ARM Ethos NPU series (Ethos-N78, Ethos-U55)
    • CEVA NeuPro-Nano (TinyML)
    • Cadence Tensilica Vision P6 (lower tier)
    • Synopsys ARC NPX6 (3,500 TOPS, 2024)

  2. More Process Nodes

    • 3nm analysis for next-gen mobile
    • 28nm for cost-sensitive IoT
    • Power/performance scaling curves

  3. Multi-Core Configurations

    • 2×, 4×, 8× IP core clusters
    • Heterogeneous configurations
    • Multi-die chiplet architectures

  4. Cost Analysis

    • Licensing fee estimates
    • Per-chip cost at volume
    • NRE (non-recurring engineering) costs
    • Total cost of ownership (TCO)

References

ARM Mali-G78

  • ARM Mali-G78 Product Page
  • ARM Mali-G78 Product Brief (2020)
  • Google Tensor SoC specifications
  • AnandTech: "Arm Announces The Mali-G78 GPU: Evolution to 24 Cores"

IP Core Comparison

  • CEVA NeuPro-M Product Brief
  • Cadence Tensilica Vision Q8 announcement (2021)
  • Synopsys ARC EV7x Product Brief (2019)
  • Industry licensing models and integration timelines

Conclusion

Successfully created a comprehensive IP core comparison tool that enables hardware architects to:

  1. Evaluate: Compare 6 licensable IP cores across key metrics
  2. Decide: Make informed SoC integration decisions
  3. Optimize: Select best IP for specific use cases
  4. Budget: Understand power, performance, and cost trade-offs

The tool provides actionable guidance for SoC design, from mobile flagship chips to automotive ADAS platforms, with realistic performance estimates and integration considerations.

Next Steps: Use this tool during the architecture phase of custom SoC design to select optimal IP licensing mix before committing to tape-out.