Horizon Robotics Journey™ 5

The new AI Inference processor
background image

Designed for AI compute

True AI performance measured in accuracy, speed and efficiency

40
Gbps
Video input
128
TOPS
INT8 peak compute
60
ms
Typical Latency
30
W
Typical power in ADAS/AD

EFFICIENT AI PACKED INTO HORIZON'S HIGHEST PERFORMANCE PROCESSOR WITH HIGH-LEVEL AD CAPABILITIES

HORIZON INNOVATIVE “BPU” AI ENGINE ARCHITECTURE

Bayes BPU: Purpose-built for Automotive Intelligence

  • - Highly parallel AI processing with multiple concurrent compute units
  • - A powerful systolic array of tensor cores for efficient CNN operations
  • - Flexible, large scale near-compute-memory for local execution
  • - High-flexibility high-bandwidth custom concurrent data bridge
  • - Data parallelism. Kernel parallelism. Computation unit parallelism
  • - Instruction parallelism. Layer parallelism. Model parallelism

(*) Tensor unit: Conv operations
Data reshape unit: Padding, reorder, Concat …
Domain specific unit: Pool, resize, warp …
Vector unit: Elem wise calc. on vector and matrices
(row-by-row, column-column or tile-by-tile)

JOURNEY 5 AUTOMOTIVE INTELLIGENCE PROCESSOR

Rich and Diverse Computing for End-to-End Autonomous driving

Deep Learning:

- 128 AI TOPS
- Highly parallel architecture
- Optimized for perf. at low batch
- Memory-near-compute /
custom data bridge
- Systolic CNN compute units
- Feature and weight compression
to save memory bandwidth

System support:

- Up to 16 Camera ingest @ 40Gbps
- 4x 4 MIPI CSI2 RX ports / 16 lanes
- MIPI TX output to display @ 20Gbps
- 2x GbE with TSN/PTP support
IEEE 1588 and IEEE 802.1
- 64b LPDDR4/x with inline ECC
- PCIe Gen3 2 lanes
- 4x CAN FD for Radar integration

Compute Diversity:

- 8x Cortex-A55 CPU
- 2x ISP @ 1.3Gp/s. HDR pixel
tone mapping
- CV engine (GDC, Pyramid,
Stitching, Optical Flow)
- 2x Vision P6 DSP for flexibility
- 4K HEVC Codec for data logging
- Hardware Crypto with Elliptic curve

Safety and reliability:

- 2x core lockstep M-Series MCU
- Safety island for system monitoring
- ASIL B SEooC
- Trustzone. TRNG. Latest Crypto
- AEC-Q100 Grade 2
- Power efficient at 25-30W typical
- 16nm FinFET
- 27x27mm FCBGA1588 / 0.65mm

EFFICIENT ORCHESTRATION OF A DIVERSE COMPUTING ENSEMBLE

Integrated, Streamlined Compute Collaboration with High Flexibility, Efficiency and Safety

STEADY PROGRESS TOWARD JOURNEY 5 SAFETY GOALS

JOURNEY 5 CYBERSECURITY

Comprehensive security measures with hardware acceleration and ARM TrustZone architecture

Trusted & dedicated resources

Secure boot, clock, compute and memory on ARM Trust
zone (TZ) architecture. Memory protected by ECC

Fully encrypted & HW-accelerated

Extensive crypto engines, Hardware TRNG and Key
generation

I/O Protected

Hardware firewall for interconnect, DDR and I/O &
peripherals; Debug (JTAG) with Authorization

OCOTP

On chip one-time programmable memory

JOURNEY 5 EVM KIT

Rich and Diverse Computing for End-to-End Autonomous driving

MATRIX® 5 INTELLIGENT VEHICLE COMPUTING CENTER

Horizon Open Reference Design Accelerates Partner’s Time to Success

MATRIX® 5 ACCELERATION CARDS

Efficient AI Computing

Let’s talk about our journey together!