Skip to main content

Horizon Journey 2 (Journey 2)

Product Overview​

Horizon Robotics was founded in 2015 by Yu Kai, former head of Baidu's Institute of Deep Learning, and is one of China's earliest startups focused on edge AI chips. The company developed its own BPU (Brain Processing Unit) compute architecture, naming each generation after a mathematician — the first Gauss, the second Bernoulli, the third Bayesian, and the fourth Nash.

The Journey 2 was officially released in August 2019 and announced as entering mass production at the World Artificial Intelligence Conference in September 2019, making it China's first automotive-grade mass-produced AI chip. The chip was designed from the outset to the AEC-Q100 automotive electronics reliability standard, operating from -40°C to 125°C, filling the gap of "zero domestic automotive-grade mass-production breakthroughs" in autonomous driving chips at the time.

The Journey 2 carries what Horizon calls "BPU 2.0" — the second-generation BPU compute architecture (i.e., Bernoulli 1.0) — using the TSMC 28nm HPC+ process in a 17×17mm BGA package, delivering over 4 TOPS of equivalent compute at a typical power of only 2W. Horizon emphasizes the "effectiveness" of its compute: typical algorithm models achieve MAC utilization >90%, and frames processed per TOPS can exceed 10× that of equal-compute GPUs.

The product serves two main lines: first, entry-level ADAS vision perception (front-view all-in-one solutions, completing detection of 24 object categories with 60 targets output per frame at <100ms latency); second, smart cockpit human-machine interaction (face recognition, gaze tracking, gesture recognition, voice). The first mass-production model was the Changan UNI-T (launched 2020-06, carrying the intelligent cockpit NPU compute platform jointly developed by Changan and Horizon), followed by the Chery Ant, SAIC, GAC Aion, and other models.

Core Specifications​

ParameterValue
ArchitectureHorizon BPU 2.0 (Bernoulli 1.0), dual-core BPU
Process NodeTSMC 28nm HPC+
CPU2 × Arm Cortex-A53
FP16 / BF16 ComputeNot disclosed (the BPU is a fixed-point integer inference acceleration unit)
INT8 Compute> 4 TOPS (equivalent compute)
FP32 ComputeNot disclosed
Memory CapacityUp to 4 GB (off-chip DDR4 / LPDDR4, 32-bit)
Memory TypeDDR4 / LPDDR4, up to 3200 MT/s
Memory BandwidthAbout 12.8 GB/s (estimated from 32-bit × 3200 MT/s)
TDP2 W (typical power)
InterconnectNot disclosed (automotive SoC, no multi-card interconnect; multi-chip solution cascading possible)
InterfaceSoC onboard (17×17mm BGA package); max input resolution 4K, dual video input support
Vision/ImageISP supports 2×HD + 1×4K; pixel-level semantic segmentation of >60 object classes
Automotive CertificationAEC-Q100; operating temperature -40°C ~ 125°C
Launch2019-08
Mass Production/AvailabilityMass production announced 2019-09; first production model Changan UNI-T launched 2020-06

⚠️ Specification uncertainty: On the CPU cores, official release materials and mainstream media report a dual-core Arm Cortex-A53, but some broker research reports state 2 × Cortex-A35, and one report's table mistakenly lists the "Journey 2" as 16nm / 5 TOPS (that parameter set actually corresponds to the Journey 3). This card follows Horizon's official statements and first-hand 2019 reporting: 28nm HPC+ / >4 TOPS / 2W. Memory bandwidth is an estimated value from bus width and data rate, not an official figure.

Key Features​

  • China's first automotive-grade mass-produced AI chip: mass production announced 2019-09, breaking the domestic autonomous driving chip mass-production gap
  • Extreme energy efficiency: >4 TOPS equivalent compute / 2W typical power, no active cooling required
  • High compute effectiveness: typical model MAC utilization >90%, frames per TOPS over 10× that of equal-compute GPUs
  • Low-bandwidth design: only 32-bit DDR memory required, reducing system power and BOM cost
  • Perception capability: latency <100ms, recognition of 60+ object classes, 200 targets detected per frame; ranging/speed-measurement error better than mainstream contemporaries
  • Dual-domain coverage: the same chip supports both ADAS front-view perception and smart cockpit multimodal interaction
  • Companion toolchain: released with the Horizon OpenExplorer (Tiangong Kaiwu) AI chip toolchain and development kit
  • Open business model: Horizon positions itself as a Tier 2, offering layered delivery from reference designs to chip + toolchain (in contrast to the Mobileye black box)

Vendor Information​

ItemDetails
CompanyBeijing Horizon Robotics Technology Co., Ltd. / Horizon Robotics (HKEX 9660.HK)
HeadquartersBeijing, China
Founded2015
Founder / CEOYu Kai (former head of Baidu's Institute of Deep Learning)
Core TechnologyProprietary BPU compute architecture + Tiangong Kaiwu (OpenExplorer) AI development platform
Websitehttps://www.horizon.auto

Use Cases​

  • ✅ Entry-level ADAS / L1–L2 assisted driving front-view all-in-one units
  • ✅ Smart cockpit DMS/OMS: face recognition, gaze tracking, gesture and voice interaction
  • ✅ Low-cost, passively cooled automotive vision perception modules
  • ✅ Aftermarket ADAS and commercial vehicle vision solutions
  • ❌ Advanced urban NOA / L3+ autonomous driving (far insufficient compute)
  • ❌ BEV + Transformer end-to-end large model deployment
  • ❌ Data center training / inference (positioned as an automotive embedded SoC)
  • Horizon Journey 3 — Same vendor's next generation, 16nm / 5 TOPS, direct successor on the Bernoulli 2.0 architecture
  • Horizon Journey 5 — Same vendor's high-compute generational leap, 128 TOPS Bayesian architecture
  • Horizon Journey 6 — Same vendor's current flagship, 7nm / 560 TOPS Nash architecture
  • Black Sesame Huashan A1000 — Contemporary Chinese automotive competitor, 58 TOPS driving-parking integrated
  • NVIDIA DRIVE Thor — International high-end automotive platform, illustrating the compute scale gap

References​