7 years ago they were using 100 TOPs.
Today they are definitely in the hundreds of TOPs and maybe up to 1,000 TOPs.
But I don't see them being much over 1k.
But no one knows for-sure.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Wiki Consumer AV - Status Tracking Thread
- Thread starter EVNow
- Start date
Today they are definitely in the hundreds of TOPs and maybe up to 1,000 TOPs.
But I don't see them being much over 1k.
But no one knows for-sure.
diplomat33
Average guy who loves autonomous vehicles
Thanks for the info. The reason I guessed 1000+ TOPS is because I know Waymo has a lot of sensor data from cameras, lidar and radar as well as a lot of cutting edge NN for perception, prediction and planning. I figured all that probably requires a lot of compute. Although I know Waymo has been working to reduce the compute of their lidar perception. At ECCV 2022, Anguelov presented their work on Range Sparse Net, a new lidar model that does not depend on range and therefore is less compute intensive.
diplomat33
Average guy who loves autonomous vehicles
"NZP will cover 60 cities across the country in March 2024; in the second quarter of 2024, it will cover more than 90% of prefecture-level cities in the country"
Looks like SuperVision on city streets is finally launching!
Looks like SuperVision on city streets is finally launching!
diplomat33
Average guy who loves autonomous vehicles
I was really just curious what Bladerskb thought about how many TOPS Waymo uses since I have not heard much discussion on that.
I agree we should not compare ME to Waymo since they have very different approaches and business models. ME's business model is selling ADAS to OEMs with the hope of eventually providing L3 and L4 products to OEMs as a way of making more money. Since consumer cars have very narrow profit margins, cost is essential. So consumer cars need a low cost, low compute solution. You cannot put the expensive sensor suite of a Waymo on a consumer car. Hence, it makes sense for ME focus on providing as much autonomy as possible with the lowest sensor suite and lowest compute possible. And incidentally, this is why Tesla is also trying to "solve FSD" with a vision-only approach. Waymo does not have these same constraints since they are doing robotaxis only. So they can put more sensors and more compute on their cars.
To your point, yes, if you have more sensors and more high res sensors, processing the data will require more compute. But I think it also really depends on the software. I am not an expert but I imagine that there are some software approaches that might be more compute intensive than others. So for example, you might have a vision NN that is not very efficient and requires more compute or another vision NN that is more efficient and requires less compute, on basically the same sensor.
And I don't think we should ignore the compute requirements of the prediction and planning stacks either. ME has talked about how there are "brute force" approaches to doing behavior prediction where the NN models every single possible path that all objects in your scene might take. These models are very compute intensive because the number of possible futures can grow very quickly. Or you might do behavior prediction by only modeling "reasonable futures" which will require less compute since the number of futures is much lower. ME proposes a different model that does not predict exact paths at all but instead only predicts intent of other vehicles and uses their RSS safety model to ensure a safe path. This model requires even less compute since it does not need to model paths for objects. My point being that different approaches to doing behavior prediction can require more or less compute. Same is true for driving policy or planning. An approach where you use 300,000 lines of code for your driving policy will require different compute than an approach that uses a NN for driving policy.
diplomat33
Average guy who loves autonomous vehicles
Xpeng's XNGP expands to cover all roads in China:
cnevpost.com
Xpeng says its XNGP assisted driving feature becomes available on all roads in China
Xpeng will begin development of a globally available highway NGP in 2024 and a globally available XNGP in 2025.
cnevpost.com
I've never watched it before. Similar to FSD Beta in some functionality.Xpeng's XNGP expands to cover all roads in China:
Xpeng says its XNGP assisted driving feature becomes available on all roads in China
Xpeng will begin development of a globally available highway NGP in 2024 and a globally available XNGP in 2025.
diplomat33
Average guy who loves autonomous vehicles
Yeah, XNGP is basically Xpeng's version of FSD beta. I would say it is probably similar in functionality and performance to V11.
diplomat33
Average guy who loves autonomous vehicles
According this, the disengagement rate of XNGP on highways is 1 per 2,000 km (1,242 miles).
First promo vid of Mobileye city streets supervision
www.bilibili.com
极氪001城市NZP，上海雨夜晚高峰实测_哔哩哔哩_bilibili
极氪 https://www.zeekrlife.com/zeekr001搭载Mobileye SuperVision™的极氪001城市NZP，上海雨夜晚高峰实测，全程0接管。轻松应对诸多现实道路场景，你值得拥有。（视频来源@极氪 ）, 视频播放量 12651、弹幕量 1、点赞数 35、投硬币枚数 4、收藏人数 14、转发人数 80, 视频作者 Mobileye官方, 作者简介 迄今为止，全球约有1.7亿辆汽车搭载Mobileye的技术。，相关视频：虽然它只有48tops的算力，也没有激光雷达，跑个高速NZP还是很轻松的！，极氪NZP暴雨体验！一镜到底无删减，拒绝充值！极氪001...
www.bilibili.com
diplomat33
Average guy who loves autonomous vehicles
First promo vid of Mobileye city streets supervision
极氪001城市NZP，上海雨夜晚高峰实测_哔哩哔哩_bilibili
极氪 https://www.zeekrlife.com/zeekr001搭载Mobileye SuperVision™的极氪001城市NZP，上海雨夜晚高峰实测，全程0接管。轻松应对诸多现实道路场景，你值得拥有。（视频来源@极氪 ）, 视频播放量 12651、弹幕量 1、点赞数 35、投硬币枚数 4、收藏人数 14、转发人数 80, 视频作者 Mobileye官方, 作者简介 迄今为止，全球约有1.7亿辆汽车搭载Mobileye的技术。，相关视频：虽然它只有48tops的算力，也没有激光雷达，跑个高速NZP还是很轻松的！，极氪NZP暴雨体验！一镜到底无删减，拒绝充值！极氪001...
Thank you so much for sharing. Nice to finally see something of in-production SV on city streets.
One thing that really stands out to me is how busy and congested Chinese cities are. That's a lot of vehicles and pedestrians often getting in the path of the vehicle. And it looked like there was a little rain too. It is not the easiest driving environment. Obviously, it is a promo video but SV does seem to do pretty good with navigating around everything. I did notice though at 1:55 mark the car stops awkwardly as it almost hit a motorcyclist.
Do you know the sensors on the car? I am just curious if the car has a forward radar or lidar helping SV avoid collisions or if it is pure vision. Thanks.
I look forward to seeing more videos of city driving as SV rolls out to more cars.
11 cameras and one forward radar.
It actually might be the smallest part of it. What really matters is the resolution of the model you are running, the complexity of the models' architecture and the FPS you are running the model at. Mobileye for example uses older architectures like CNN that are a-lot more efficient, cropped out resolutions, very low FPS and then also traditional computer vision that are very very very efficient.
However running SOTA complex architecture models like Waymo does, at full resolution and very high FPS requires very high compute.
Are you serious ? There is more congestion with traffic and pedestrians (and construction) in a suburb of Seattle like Bellevue than what you see in that video !
But, yes, thanks for the video @Bladerskb . What I’d like to really see are independent longer videos by consumers like we see here with FSD.
Sorry, but I do not understand technical terms like "SOTA complex architecture"
"running SOTA complex architecture models like Waymo does" means literary nothing. Is it considered to be."SOTA" because it has many sensors at high res or SOTA because it's complex? Or is is SOTA and complex because it isn't optimised or what's the idea?I stand by my previous statement. Ie Many high res inputs into the nets is what's driving compute cost more than anything else. Processing 25+ high res feeds is very costly even if it's just merging and getting the BEV afaik.
Last edited:
You are both saying the same things but in different ways.
Sensor processing is the most computationally expensive. Well with high resolution sensors as well the type of architecture and software you need to process the volume of data coming in realtime plays a role in how much compute you require.
You can always throw more compute at a problem but now you're consuming more power which is limited on an autonomous vehicle. Everyone is finding ways to optimize their architecture to do more with less compute.
Sensor processing is the most computationally expensive. Well with high resolution sensors as well the type of architecture and software you need to process the volume of data coming in realtime plays a role in how much compute you require.
You can always throw more compute at a problem but now you're consuming more power which is limited on an autonomous vehicle. Everyone is finding ways to optimize their architecture to do more with less compute.
diplomat33
Average guy who loves autonomous vehicles
Zeekr rolled out NZP (Mobileye Supervision) in 36 cities for both Zeekr 001 and 009:
And here is video of Zeekr's LCC+ feature showcasing its ability to avoid road obstacles. My understanding is that LCC+ is an enhanced lane centering feature that can also steer around obstacles in the lane.
And here is video of Zeekr's LCC+ feature showcasing its ability to avoid road obstacles. My understanding is that LCC+ is an enhanced lane centering feature that can also steer around obstacles in the lane.
Is this highway only or urban too i.e. is this like FSD or NOA ?
Point-to-point
automated navigation
& adaptive cruise
control
diplomat33
Average guy who loves autonomous vehicles
NZP is now rolling out to city streets. So it is now urban too. So it is point to point, automated navigation on both highway and city streets like FSD beta.
Still no independent long drive video .... that is what we need to evaluate it vs FSD.
diplomat33
Average guy who loves autonomous vehicles
Not yet. It just rolled out. I am sure we will get some independent long drive videos soon.
Similar threads
