
Journey 6 Automotive Chip: A Walk Through the Deployment Pipeline
As someone who knows nothing about chips, I tried walking through the entire model deployment chain for Journey 6. This thing isn't meant for individual hobbyists—whether it's worth using depends on whether you're doing mass-production smart driving. For OEMs and Tier1s doing L2+ mass production and pinching costs, it's an option very much worth a serious look; for individual developers who want to buy a board and run large models at home, this path basically doesn't work.
Journey 6, abbreviated in the industry as J6, is Horizon Robotics' in-vehicle computing chip for intelligent driving. Note it's a family, not a single chip. By compute power from low to high it's split into six models: J6B, J6L, J6E, J6M, J6H, J6P. I didn't get it at first—took me two passes through the docs to sort out. The lower ones handle active safety and basic assistance; the higher J6E and J6M carry L2+, and above that is city smart driving. The numbers given in the docs: J6E is 80 TOPS, 100K DMIPS; J6M is 128 TOPS, 137K DMIPS. TOPS is trillions of operations per second, measuring AI compute; DMIPS is the CPU's integer computing capability. You have to look at both numbers together to roughly know what tier of work this chip can do.
There's also a term called "four chips in one," meaning BPU, CPU, GPU, and MCU processing units are integrated onto the same chip. BPU is Horizon's own AI acceleration unit. The benefit of integration is saving board space and, arguably, power.
I don't have a dev board, and I'm not going to pretend I do. What I did was fairly basic. The official developer community has a set of OE-Skills that claims to support agents like Codex, so I set up a project environment, wanting to first sort out the deployment chain. The so-called deployment chain is the series of steps to get a trained model running on the chip: quantization, compilation, graph optimization, on-board validation.
The night I set up the environment, it took me about two-plus hours. The docs are in Chinese, which is friendlier than a lot of chip vendors. But I had to modify the toolchain install script three times before it passed—once for the Python version, once because there was a space in the path. These pitfalls are trivial, and newbies die on them most easily. After installing, opening that environment, the interface looks like a task panel: you write out the steps you want to do clearly, it lists them for you, and some can directly generate commands. After hooking up the agent, part of the configuration it can auto-fill following the docs, saving me the effort of flipping through the manual page by page. In the past, chip docs like this ran to hundreds of pages, basically enough to scare newbies off.
There were plenty of places I got stuck too. The first hurdle is model selection—I stared at the differences table for the six models for a long time; which one can run how big a model, which supports how many camera channels, you have to check item by item. The second hurdle is toolchain errors—when versions don't match, the prompts are very vague. I don't have real hardware here, so "running through" only got as far as compilation and graph optimization; performance wasn't actually tested, and I won't brag about that.
I want to say a bit more about cost and power. According to a Nikkei report, J6 based on reference design comparison is about 40% cheaper than the previous generation Journey 5, using TSMC 7nm. BYD's God's Eye C released in February 2025 uses this series. What does 40% cheaper mean—it means cars in the hundred-thousand-yuan range can also get decent assisted driving. That push to the industry is more real than the compute numbers. On power, I recall Horizon mentioning when talking about Journey 5 that a 30-watt compute platform versus a 300-watt platform can differ by 50 to over 100 kilometers of range at the vehicle level. EV batteries are only so big; chip power savings directly convert to mileage.
I did AI underwriting for a few years, and I can't shake the habit of how I look at chips. How fast the compute is comes second; I care more about whether, when it makes decisions for smart driving, it can leave a reviewable record. In a car it's harsher than an insurance policy—if something happens and you can't explain why it judged the way it did at the time, you won't pass the regulatory gate. This point can't be seen just from specs; you have to wait for real vehicles and Tier1 deployment cases.
Let me also spell out the downsides. Individuals simply can't get their hands on the hardware, only look at specs and others' tests. The six model names are unfriendly to newbies—the letters don't correspond to compute, you have to memorize them. Actual performance you have to trust the OEMs' and Tier1s' test data; you can't verify it yourself.
If you're also doing model selection, my advice is don't rush to pick the chip first—first figure out clearly how many models you want to run, how many camera channels, and which tier your budget is capped at, then go back and match against the six models. If they don't match, everything afterward is rework.
Physix Frontier