Recently, I have noticed that there are more and more AI development boards available. It can even be said that if future development boards do not have AI capabilities, one might reconsider whether to purchase them; it’s that exaggerated.The development of AI has also been quite rapid. Previously, I thought of it as a form of artificial intelligence that was somewhat foolish, like a cyber seamstress. However, since the emergence of DeepSeek, it feels like AI’s intelligence has made a significant leap forward.There have been breakthroughs in various fields, and I recently saw that Huawei proposed a “Mathematical Theorem Proving Agent” (LEGO-Prover) in their 2024 annual report, which utilizes theorem provers, theorem evolution engines, and dynamic skill libraries for interaction and collaboration to achieveautomatic theorem proving. I find this quite remarkable.
Combined with the fact that all major companies are focusing on this area, I believe AI is the path we need to take now, and it should be the right one..As for development boards, to be honest, due to hardware and size limitations, I don’t think there have been any qualitative changes, but there are many powerful development boards available.However, the application of development boards in AI differs from scenarios that require significant computing power, such as large language models and text-to-image generation.Applications like image recognition and speech recognition are more prevalent.Below, I will recommend a few AI development boards.First, I recommend theESP32S3 development board.
Although its performance is quite weak and it lacks dedicated AI computing modules, there have been many AI-related projects recently, mainly focused on the voice assistant field, combining large language models and speech recognition models to create fun and portable AI scenarios.
If you just want to experience the charm of AI and have some hands-on ability, you can buy one to play with; it will be more enjoyable than using a smartphone.Next is theSeeed Grove Vision AI V2
It features a dual-core Cortex-M55 CPU and ARM Ethos-U55 NPU, has a low entry barrier, and allows for easy image recognition. It is very simple to connect and use with Arduino or ESP32..
The chips from Canaan Creative are also impressive. Previously, theK210 was widely used for image recognition and was quite popular.
Now they also have theK230 and K230D, which have stronger computing power, with an equivalent computing power of 6 TOPS. This level of computing power is sufficient for most image recognition scenarios and is also a popular solution in current electric competitions.
Next, we have the Rockchip camp, with RK3566, RK3568, and RK3562 all featuring Rockchip’s1 TOPS NPU, which can also perform various simple model inferences.
The SG2002 also has 1 TOPS of computing power and is RISC-V based, allowing for a choice between RISC-V and ARM cores, making it quite versatile. Many development boards currently use this chip.
Next is the JIT Space K1, which is a RISC-V chip with2 TOPS of computing power.
It also features a CPU that integrates general-purpose AI computing power, which should not be underestimated. This may become a key focus for future development boards in AI. Currently, many development boards are adopting the JIT Space K1.
Moving on, we reach the 6 TOPS computing power range. This level of computing power is quite familiar to us, specifically the NPU used in various Rockchip models like RK3576, RK3588, and RK3588S, which can deploy models up to 7B.
8 TOPS of computing power is primarily used in Huawei’s Ascend 310 SoC, which also has a 20 TOPS option available.
Huawei’sAtlas development board and Orange Pi AI Pro development board use this chip.
The Digua Robot RDK X5 uses the Sunrise 5 chip, featuring 8 Cortex-A55 cores, with a BPU computing power of10 TOPS, performing well in AI and providing a rapid development platform.
The Cube Pi 3 is a Qualcomm open-source development board that I recently acquired, with a computing power of12 TOPS, which is also quite good.
The Raspberry Pi 5 can also play with AI by adding the AI Kit module, offering options for13 TOPS and26 TOPS, but users have reported that the computing power is somewhat inflated, and the actual experience is average.
The Radxa Orion O6 uses the CD8180 SoC, with a computing power of30 TOPS. This chip not only has strong AI computing power but also boasts a powerful CPU and GPU, far exceeding the RK3588. Unfortunately, I have not yet obtained a board to test its capabilities.
The SG2300X SoC, also known as BM1684X, can achieve32 TOPS or 16 TFLOPS (FP16/BF16).
This chip is also used in some development boards, primarily for AI recognition in video surveillance, capable of simultaneous recognition of 32 channels, and can handle various large models, including text-to-image generation and large language models.Moving up, we enter the realm of NVIDIA, from Jetson Orin Nano to Jetson Orin Nano Super, and for even stronger performance, there is the275 TOPS Jetson AGX Orin.
This is basically the pinnacle of AI performance in development boards, but not the pinnacle of AI devices.If we talk about strong performance, most of what we can see and access is concentrated in the field of intelligent driving.
For example, NVIDIA’sDRIVE Thor has a computing power of2000 TOPS, which has not yet officially been deployed in vehicles, but I see that many car manufacturers are preparing to install it in new cars.
Can we go even further?Yes, beyond this is the world of various AI servers and supercomputers. For example, Huawei’sAtlas 800 inference server features two Kunpeng 920 CPUs, with a computing power of1120 TOPS, with a single Kunpeng 920 providing 560 TOPS. To achieve strong computing power, one simply needs to stack more cores.
Huawei has also launched theAtlas 900 AI cluster andAtlas 900 A2 PoD cluster basic unit, with computing power that can be increased based on the number of basic units needed.
Each basic unit has32 Kunpeng 920 CPUs, achieving a computing power of17920 TOPS, with 256 DDR4 memory slots, each supporting 32GB or 64GB. The maximum power consumption is 50.5KW, with liquid cooling.
Huawei also has solutions forurban AI hubs andnational regional AI computing centers..
It can be said that the strength you desire depends on the size of the venue and whether you can afford the equipment and electricity costs.
NVIDIA also has similar devices, stacking cards and cores. A single core’s computing power is even stronger than Huawei’s.
Overall, currently, AI development boards are limited by computing power, primarily focusing on inference and application aspects. This aligns with their size and positioning. However, even so, in high-end scenarios for AI development boards, currently, only NVIDIA stands out. Huawei has powerful chips, but currently only offers computing cards, not development boards.
In lower computing power scenarios, there are many players, currently mainly concentrated within30 TOPS, but even so, there is a loose ecosystem, and using them generally requires model conversion, and the actual usage also depends on the design of the NPU and AI units, which may lead toinflated computing power claims, or chips that are particularly optimized for specific fields, being very strong in those specific areas.We can also observe that in the AI field, development boards with RISC-V cores are slowly emerging. I have always believed that IoT and AI are the breakthrough points for RISC-V, but currently, there are still not enough standout products. I look forward to more exciting chips and products emerging.
Our public account has launched AI intelligent Q&A
We recommend development boards, search for development board parameters, and other fun experiences
Development board selection website, everyone is welcome to visit, although it is not yet very complete, it has already recorded about 500 development boards. If you are interested or have needs, you can take a look at our selection website. You can also share any good suggestions in the group or leave us a message.
Our website domain is:findboard.cn
（We have not yet adapted for mobile, it is recommended to use a computer browser for access）