Best edge boards for robotic // Tier List

Ajouté : 2026-05-05

[00:00:00]Hi everyone. In my last video, I told you about different boards for VLMs and LLMs. And in this video, I want to rate boards for robotics.

[00:00:12]And to do this, first I will split all boards in four different categories.

[00:00:19]First of them, the board can run only classical computer vision models like detection, classification, regression, and so on.

[00:00:30]Boards of these types, they are great and actually they can cover 80% of modern robotics, in my opinion. The second level, it's boards that can support LLMs and VLMs. Actually, you saw a lot of them in my previous video, but for robots, it's better than like basic model support.

[00:00:51]And actually, with VLM support, you already can build some type of VLA models. Third level, it's boards that can support not only detection, classification, regression, VLM, LLM, but also stereo depth estimation model.

[00:01:07]Why it's important? Because actually, robots, they are manipulating with space around us. And this support of stereo depth estimation model, it's pretty important when you want to estimate your object precisely. Because if you need to grasp, sometimes it's not enough just one D image. Yes, sometimes you can work with images from three cameras from VLAs models, but it's not super accurate sometimes. And sometimes, when you work especially with high speed, you need depth image. Yes, you can take classical depth estimation camera and work with them. But it's always nice when you can estimate stereo depths with other on your edge with good with good quality. Fourth category, it's fourth category, VLA models. Actually, it's not a lot of boards which can support VLA. And we will discuss this when we will go deeper into this. So, let's start. First level, basic models.

[00:02:17]And the worst board, Jetson DLA. Not the Jetson itself, but actually Jetson DLA is one of the worst way to inference your models. For Jetson, you can run classification and on base of classification, you can trade some regression, of course. But for Jetson DLA, of course, there will be no support of LLMs, VLMs.

[00:02:42]Maybe you will support some old detection models. And because of this difference between Jetson Nano, Jetson and NX, it's not so dramatically different. Because from Jetson, you need just GPU, and DLA it's >> [snorts] >> not working.

[00:02:59]Next level in level one, VeriSilicon. They are actually great. I did a lot of review on my channel. You can check UB boards, you can check others board and other board. But with them, you just can't run any model, VLM, LLM, stereo depths, and so on. And pretty nice board, but not here. Next level, Texas Instruments. In my opinion, they are a little bit better than VeriSilicon because better support, better model, better NPU available, but not as good for the next level. And the best board for this level, in my opinion, it's MemoryX. Because actually, they support great. It's like one button to export. But again, no LLM, VLM, and so on. But if I would choose a board for this level, definitely MemoryX one of the best options here. Let's go next. Level two, LLMs, VLMs.

[00:04:04]If you want more detail, you can go to my previous video. But here, I want specify a few boards that's are not on the level one uh already, but still not on level three and level four. First, start from Hailo 10.

[00:04:24]I didn't test it yet. And I heard that it's not the fastest board, but it can inference some LLMs and VLMs.

[00:04:34]Next level, Accelera. They are much faster, but also some basic LLMs. I think they supported one VLM recently.

[00:04:43]So, yes, on this level.

[00:04:46]Uh next board, Rockchip. Actually, great board for small models, but again, not possible to run any bigger model. So, Rockchip is stopping here.

[00:04:59]And the next level is mm IMD NPU. Because I were not able to run any stereo depth estimation model and VLA model on them.

[00:05:12]Let's go on the level three.

[00:05:15]And here, the situation is pretty interesting because different boards support different models. And maybe I need to mention that Hailo 8 can support some old old models. I don't know about Hailo 10, but maybe theoretically, Hailo 10 will be at this category. But according to their official repository, only one super old model. So, like somewhere between level two and level three.

[00:05:46]Also, I need to mention a pretty interesting uh point here. It's uh RDK-X5, which I had on my channel. And actually, I recently have this camera 3D depth camera from Loopar Robotics, which use the same RDK-5. But because they trained specific model for this uh stereo camera, this out-of-the-box uh it's not out-of-the-box model.

[00:06:15]And when it's trained specifically here in this environment, it perform much much better than the original model. But you don't have like the full access to this model. And because of this, it's not the full experience of running this uh stereo depth estimation on this level three.

[00:06:33]>> [snorts] >> And because of this, in my opinion, RDK-X5, it's like on the border.

[00:06:39]What next?

[00:06:40]I want to mention a few boards that are here, but they support like partial.

[00:06:50]It's Axent, Huawei. There was a video about Orange Pi on my channel where I where I run some models and some other didn't work. Uh the same with Safran NPU. And uh the same with SpaceMeets.

[00:07:07]Some models were able to run and some not.

[00:07:11]I need to mention that uh there is another company which I forgot in my previous uh overview. It's Accelera. And I expect that they able to run some models. But with them, it's I don't have access to a new board. And it's hard to collect information for them. So, I put them on this uh level. And of course, I need to mention Qualcomm and specifically solution from Oak, where they created specific board for uh this Oak, but also on Qualcomm, they were successfully the they succeed to run a few existing stereo depth estimation model, and not the simplest one. So, in my opinion, in this specific category, uh Qualcomm is leading. And uh it actually can run a lot of models. Here, I mentioned specifically Qualcomm NPU.

[00:08:10]Okay, let's go deeper on the level four, VLA.

[00:08:15]And here, of course, we have just a few competitors because you can run your models only when you have a full PyTorch support for VLAs. And right now, in my opinion, it's just three vendor, IMD, Intel, and of course, Nvidia. So, the board from them, it's the board from Intels, which pretty small like Intel NUC classes boards. Of course, it's some small board from IMD, but they are not fully functional, and uh they still slower than other board. And of course, it's Jetsons. So, only three of them can fully support PyTorch. And because of this, they can fully run VLA models. And up to date, it's not possible to run all these VLA models on any other inference devices, sadly.

[00:09:17]There are some work from ExecuTorch, which is supported by GPU on Qualcomm.

[00:09:23]And when they fully support this, I hope that ExecuTorch with Qualcomm also will be an option to run such models, but not mm working yet. I checked on some issues on the forum, and people actually were not able to run any full VLA model.

[00:09:43]And the same issue with OpenVINO. And of course, it's the same issue with ONNX and uh everything that is working on base ONNX.

[00:09:53]So, level four, in my opinion, it's Jetson, Intel, AMD.

[00:10:00]And this is the current situation. So, if you need like the boards that fully support any models for robotics up-to-date, it's just Jetson and maybe some Intel board with GPU or AMD. If you are okay with like some stereo depth estimation, some LLMs, VLMs, it's a few other boards. Probably Qualcomm is the best here of core and a few boards with a lower support. And of course, LLM, VLM, you have much more models available. And if you need just simple robots that will detect some objects and do some easy actions, for example, like classification and so on, then you have access to all existent boards in the market.

[00:10:51]Hope this overview was helpful for you and I meet you in the next video.