New Energy Vehicles (NEVs) in China

More to think about ADAS. For BYD, the guy in charge of it is Hanbing. He took over in May after the old head resigned and that team of over 500 engineers joined his electronic integration team. This team takes over ADAS now

比亚迪智能驾驶研发负责人王欢已经离职，其负责的智能驾驶开发部被分拆整合，该部门超500名员工，大部分被分流进韩冰负责的电子集成部。此次调整完成后，电子集成部总监韩冰已经成为比亚迪规划院的智能驾驶研发负责人。与此同时，韩冰还在同步筹备比亚迪的智能驾驶芯片设计团队。

Click to expand...

he presented back in June and showed the following

在今天的 2023 北京智源大会上，比亚迪规划院院长助理兼电子集成部总监韩冰谈到了一些比亚迪智能驾驶的数据情况。

关于数据闭环的系统建设、数据积累方面：比亚迪现有智能驾驶研发车队约 300 辆、数据规模约 150 PB 的历史数据，并且每天新增 1 PB 数据。（1 PB = 1024 TB）；自动标注比例超过 95%。

图像感知数据 45 亿、Lidar 感知数据 15 亿。

预计今年还会累计有 6 亿公里的数据，并在未来通过采集车辆和量产车，完成指数级的数据增长。

感知模型的开发已经做到了 100% 的数据驱动，多相机 BEV 模型处于研发中，计划今年量产。

车载端大算力平台假设方面：

这边的重点是，智驾操作系统 BOS、域控制器，比亚迪自研；核心的芯片采用的是供应商，508 TOPS 算力平台，显然采用的是英伟达 Orin-X。

下一步的应用：

首先是完成数据驱动的 Occupancy & Object 检测，可以复用过去采集的 150 PB 数据；占用网络上车，无需激光雷达即可解决地面凸起物、路牙等通用障碍物检测，实现城市端到端的智驾体验。

Click to expand...

also shown here, look at that last slide, U8 not only has 508 TOPS but also 456k DMIPS of computation, that is one powerful platform.

算力平台，只提了车端的 508 TOPS，显然是英伟达 Orin-X，但韩总全程强调自主可控，英伟达 Logo 都没出现。云端训练集群没有介绍。

Click to expand...

But note this, Nvidia's chip is standalone plug in and can be swapped out. BYD controls the entire process. Nvidia or horizon robotics aren't used for algo or computation platform
BYD doesn't even show Nvidia Logo

- 底层软件，码得很完整。 - 算法的展望，接下来要做占用网络、不依赖激光雷达和高精地图，基本上也是头部都在做的事。

Click to expand...

lower level software (OS and interaction with sensors and chips hardware) are complete, so just working on algo.
- next step is not use Lidar or high precision map

韩冰说，比亚迪感知模型的开发已经做到了 100% 的数据驱动，并同时研发了拥有多相机的 BEV 模型，计划是今年可以做到量产。BEV 模型结合比亚迪的易四方平台，可以研发出一些具有特色的高级辅助驾驶功能。

Click to expand...

BYD's sensory model is 100% data driven with multi-camera BEV model. BEV model is integrated with Yangwang's 易四方platform

韩冰还介绍了比亚迪的大模型可以用于针织系统的自动标注。比亚迪围绕相机和激光雷达结合时序建立了 4D 障碍物自动标注的系统，该系统通过历史数据的挖掘和计算可以生成一批具有真值的数据，为算法研究提供数据支持，节省人力成本。


其优势大概有四点：第一是以激光雷达的数据为核心，精度更高；第二是结合了时序的 4D 真值，感知范围更广；第三是数据驱动，以小数据驱动，再快速迭代；第四是可以迅速扩展新障碍物的类别。

Click to expand...

BYD's LLM revolves around camera & Lidar constructing 4D construction of obstacles and targets. Lidar is more accurate/precise, this approach is 4D and data driven & can more easily identify obstacles

同时整个智驾的软硬件平台，其开发都是基于模块化的分工，这样既保证可控，也可以跟外部供应商进行并行的协同开发。

好处是产品更加可靠同时降低研发成本，优质的代码和模块运用到量产产品中，这样的分工可以实现高效的合作。
这某种程度上反映了比亚迪当前的智驾研发策略：底层软硬件自研可控，上层算法模块与应用可以由供应商提供，加快量产

Click to expand...

BYD approach is to develop the OS, domain controller & software/hardeware integration. And then outside suppliers supply the algo. they are working with Momenta, but there could be others. They are using Orin initially, but they can use baidu chip, horizon journey 5, black sesame or develop their own ahead

sounds like they consider their IP to be the most important and other things like sensors, ai chips & algo are plug and play

more details on Horizon Journey 5 chip, supposedly using TSMC 16nm process

128 TOPS, 30W consumed
60ms ultra-low latency perception capability.
Through improved tool chain & compiler optimization, J5 image rate up to 1531 frames/s
neural network computation of various top level models & operators supporting transformer model of BEV sensing, BEV sensing algo
can quickly and accurately process massive camera data and complete visual preprocessing

single chip support 16-channel camera perception calculation, multi-type sensor fusion perception including MMW radar & Lidar

2Gbps ethernet comm interface -> sub-micro multi-sensor synchronization & improved RT computing efficiency

Horizon built a chip development toolkit, toolchain, SDK & domain control reference design

Journey 5 first mass produced car model is L7 & L8 AD Pro version

Also got models in development with BYD, GAC Aion, SAIC & FAW + many more

Ecosystem partner of J5 including pony.ai, QCraft (轻舟智航), holomatic (禾多科技), Minieye (佑驾创新) and others. So these are all providing ADAS solutions.

According to
Journey 5 cost $200 per chip or 1400 RMB?

地平线征程5芯片的售价为200美元/片

Click to expand...

In comparison, Nvidia Orin cost range from $399 for 70 TOPS version to $1599 for 275 TOPS version

Looks like you can also get Nvidia Orin nano for $199 & $299, but they are for 20 and 40 TOPS each

What about Black sesame?
the are really cheap

按照市场份额来看，黑芝麻智能在中国市场份额只有5.2%，却排名第三，仅次于英伟达、地平线之后。

位于行业第三，依然与行业龙头有不小的差距，黑芝麻科技在招股书中说道，旗舰A1000系列SoC的总出货量为2.5万片。作为对比，地平线征程芯片出货量突破300万片,针对L2+的征程5出货量超10万。

Click to expand...

so far Black Sesame's flagship A1000 SoC only sold 25k vs 3 million from Horizon's Journey series and > 100k for just J5
5.2% market share

中算力芯片，黑芝麻固然有华山二号A1000，算力为58TOPS，可谓瞄准了国产厂商的空白区，且售价仅100美元，但目前还未正式量产上车，市场表现有待验证。该市场还要面对英伟达、TI等成熟选手的竞争。

Click to expand...

It's marketing A1000 at just $100, capable of 58TOPS

Black sesame's great pitch is its cost

黑芝麻智能华山二号 A1000 芯片是完成所有车规级认证且较有性价
比优势的量产芯片。江淮汽车旗下思皓品牌的多款量产车型、东风集
团旗下东风乘用车首款纯电轿车和首款纯电 SUV 都将基于华山二号
A1000 系列芯片，打造行泊一体智驶域控制器平台。黑芝麻的行泊
一体方案目前已经能够实现支持 10V（摄像头）NOA 功能情况下，
域控制器 BOM 成本（bill of material，成本中的原材料部分）控制在
3000 元人民币以内，支持 50-100T 物理算力。黑芝麻的低成本主要
源于域控制器本身和其自研芯片两大因素。黑芝麻采用的是基于单
SoC 芯片的中低算力行泊一体域控方案且其华山二号 A1000L 和
A1000 芯片在性能或是成本上都可以适配行泊一体方案的需求。

Click to expand...

It can control cost for a computation platform/domain controller w/ 50-100T computation power to under 3000 RMB

此外，基于黑芝麻智能SoC的高阶NOA行泊一体域控器，能够支持100TOPS以上物理算力，并能够将BOM成本控制在3000元人民币以内，帮助车企解决成本压力。

联友智连智能驾驶产品负责人柯智皓表示：“智能驾驶产品需要快速准确地处理传感器数据和其他任务，以实现高精度的感知和决策。黑芝麻智能的高性能SoC解决方案满足了这些需求，兼顾能效和智驾体验，为联友智连打造出高品质、高性能的汽车智能驾驶产品提供了强有力的支持。相信我们双方的合作将为业界带来更多创新的智驾产品。”

黑芝麻智能产品副总裁丁丁表示：“在联友智连和黑芝麻智能的通力合作下，YDU2.0 Pro高阶行泊一体域控制器实现了让智能驾驶系统看得更‘真’、看得更‘懂’，同时还实现了单颗华山二号A1000芯片成功接入11路摄像头，这在本土芯片是首创。黑芝麻智能芯片的效能和高性价优势再一次得到验证。黑芝麻智能将与联友智连继续深入合作，赋能安全、高效、环保的智驾生活。”

Click to expand...

high level NOA solution with their domain controller support 100TOPS+ for < 3000 RMB. Can connect to 11 cameras

QCraft's story of achieve L2+ ADAS in 500 days with Journey 5

Journey 5's price to quality ratio is discussed in this article. Instead of going with Nvidia chip for L4, QCraft decided to work with J5. Cost is several thousand RMB cheaper (makes sense since J5 is about 1500 RMB, top end Orin-X is about 10k RMB (although my guess in practice, you can get it for cheaper than that)

而且在成本上，在相同算力下，征程5芯片性价比更优秀，相较于原有芯片，价格降低了几千块。

另一方面，选择征程5相比于传统的L2辅助驾驶方案，单芯片128TOPS的算力较为充裕，算法有深挖空间。

Click to expand...

BEV (Bird’s Eye View) model

高算力也让BEV等大模型的算法落地成为现实。侯聪向HiEV透露，现在现在能够用于模型训练的BEV数据已经超过了1亿帧，其中99%的数据由大模型自动标注，可以稳定确保车端模型效果。

其次，基于征程5的大算力平台，轻舟智航提出了基于视觉多任务的行泊一体BEV感知模型，可提升算力利用率并保证行车和泊车功能切换。轻舟的BEV方案采用多相机特征和时序融合的方式，生成速度和方向，解决道路语义识别、跨相机遮挡等非BEV方案模型所不具备的能力。

Click to expand...

> 100m BEV frame of data & 99% data automatically annotated by large model
it's BEV model uses multi-camera & time fusion to generate speed and direction, solving road semantic recognition, cross-camera occlusion and other capabilities that non-BEV solution models do not have.

上汽集团、长城汽车、江汽集团、理想汽车、岚图汽车、长安汽车、比亚迪、哪吒汽车等。

Click to expand...

J5 customers include SAIC, GWM, JAC, Li Auto, Voyah, ChangAn, BYD & Neta. so basically everyone

9月下旬，轻舟宣布已成功获得数家头部车企量产定点合作，轻舟乘风将在明年上半年量产上车。顺利定点也和轻舟是首家基于地平线征程5智驾方案供应商密不可分。

Click to expand...

by end of september, QCraft already agreed to cooperation with several mass produced model. To be land in PVs by first half of next year

• 轻舟乘风Air版（7V1R基于环视相机的视觉方案）：可实现高速NOA+L2全功能；
• 轻舟乘风Pro版（11V5R视觉为主的方案）：可拓展城市NOA功能；
• 轻舟乘风Max版（11V5R+1颗激光雷达）：支持轻地图模式。

这三个版本分别瞄准10万级、20万级以及20万级以上的智驾车。

Click to expand...

Developed 3 versions of their algo
1 is QCraft Air - 7 camera, 1 radar -> can achieving highway NOA & L2
2nd is QCraft Pro - 11 camera, 5 radar -> can achieve city NOA
3rd is QCraft Max - 11 camer, 5 radar, 1 Lidar -> support achieve map model
for 100k, 200k & 200k+ RMB smart cars

I'm impressed with what they are able to do

BYD U8 supposedly has 30k orders, although I don't think that's firm orders

AGV that can carry 55t payload, that has its own Lidar has rolled off production line at Xian. Also explosion proof

One obvious move recently is the proliferation of cheap ADAS solution using domestic supply chain.

I've talked about Haomo's 3000, 5000, 8000 RMB solution a while back. Well, they also work with Horizon Journey series

here they (毫末智行) talk about partnering up with "映驰科技" (enjoy move)

Using single Journey 3 chip for the low end ADAS, this is the low end ADAS package that can do parking and simple ADAS feature

100-200k is 50% of China's PV market, this is where light weight parking solution comes to play. So a lot of these ADAS tech suppliers are developing J3 powered solutions

Aside from J3, Horizon J5 & Black sesame A1000 can be used to achieve NOA and parking

宏景智驾 (hongjing drive) has already achieved Journey 3 powered drive/park solution able to lane keeping, auto parking, remote parking

福瑞泰克 (autofree tech) also works with Horizon robotics

I think the point is that it doesn't really matter who is more profitable right now. The trend is that we are at an inflection point and BYD may overtake Tesla in the near future.

supercat said:

I think the point is that it doesn't really matter who is more profitable right now. The trend is that we are at an inflection point and BYD may overtake Tesla in the near future.

https://twitter.com/i/web/status/1715770644435292433

Please, Log in or Register to view URLs content!

Click to expand...

If you add the semiconductor and battery subsidiaries (vertical integration) then BYD profit as a whole is greater than Tesla.

BYD Seal undercutting model 3 and Japanese legacy options like accord in APAC region. While pricing of seal is not that competitive in EU, they are very good in APAC region

byd bringing tang, seal, yuan+ & dolphin to Panama

Ranking of NEV Model Sales Volume by Price Range, Jan-Sep 2023
In the price range below 100k RMB, Wuling Mini ranks 1st with 169.3K units, followed closely by BYD Seagull with 113.3k units.
In the 100k to 200k RMB price range, BYD dominates with 8 out of the top 10 sales positions. Quite impressive.
In the 200k to 300k RMB price range, Tesla ModelY and Model3 occupy the top two positions, and BYD has 5 models on the sales ranking list.
In the 300k to 400k RMB price range, LiAuto L7 and L8 take the top two positions, with Tesla ModelY ranking 3rd with 58.7k units.
In the 400k to 500k RMB price range, LiAuto L9 secures the top position with sales reaching 74.0k units. #Denza D9 follows in 2nd place with sales of 52.2k units, and NIO ES6 ranks 3rd with sales of 38.2k units.
In the above 500k RMB market, MercedesBenz E-Class PHEV takes the lead, with NIO ES8 closely following.