HabrMay 15

Axera AX650N: архитектура Edge ML SoC под CNN, LLM и VLM

Большинство задач современной робототехники так или иначе завязаны на нейронных сетях: детекция объектов, оценка глубины, локализация, планирование. Всё это ресурсоёмко, и вопрос выбора компактного вычислителя (достаточно часто алгоритмы должны работать локально) встает довольно остро. На практике выбор сводится к трём классам устройств: NVIDIA Jetson , внешний ускоритель (один из самых популярных — Hailo) и китайский (не всегда, конечно, но в современных реалиях обычно китайский) SoC с интегрированным NPU. В этой статье я рассмотрю представителя третьего класса — Axera AX650N , а NVIDIA Jetson будет использоваться для сравнения, так как это единственное массовое edge-решение с универсальными вычислительными ядрами (CUDA) . Это первая часть цикла. Здесь я разберу аппаратную архитектуру самого AX650N — CPU, NPU, DSP, ISP, память — и поделюсь результатами первых тестов: YOLO, Depth Anything, SuperPoint и мультимодальный Qwen3. Подробные бенчмарки и сравнения — во второй части. Я тестировал AX650N в рамках готового устройства от Sipeed — Maix4 Hat . Он состоит из двух частей: SoM , на котором расположены SoC и 8 GB RAM (2x4 GB, так как у AX650N два отдельных DDR-контроллера) , и baseboard от Sipeed с минимальным количеством интерфейсов. Скромность интерфейсов объясняется просто: baseboard — это HAT для Raspberry Pi 5, подключающийся по PCIe 2.0. В такой конфигурации AX650N работает как внешний ML-ускоритель, аналогично Hailo . В рамках этой и последующих статей я буду использовать Maix4 Hat как самостоятельный микрокомпьютер.

https://habr.com/ru/articles/1035776/

#Axera #NPU #VLM #llm #embedded #cnn #нейросети #sipeed #электроника #компьютерное_зрение

Axera AX650N: архитектура Edge ML SoC под CNN, LLM и VLM

Maix4 HAT Большинство задач современной робототехники так или иначе завязаны на нейронных сетях: детекция объектов, оценка глубины, локализация, планирование. Всё это ресурсоёмко, и вопрос выбора...

Хабр
LinuxGizmos.com [Unofficial]May 10

Sipeed launches K3 Pico-ITX and CoM260 boards with SpacemiT RISC-V SoC

https://fed.brid.gy/r/https://linuxgizmos.com/sipeed-launches-k3-pico-itx-and-com260-boards-with-spacemit-k3-soc/

cnx-software.com [Unofficial]May 11

RVA23-compliant K3 Pico-ITX SBC and K3-CoM260 SoM feature SpacemiT K3 octa-core RISC-V AI SoC, up to 32GB RAM, 256GB UFS

https://fed.brid.gy/r/https://www.cnx-software.com/2026/05/11/rva23-pico-itx-sbc-spacemit-k3-octa-core-risc-v-ai-soc-up-to-32gb-ram-256gb-ufs/

RVA23-compliant K3 Pico-ITX SBC and K3-CoM260 SoM feature SpacemiT K3 octa-core RISC-V AI SoC, up to 32GB RAM, 256GB UFS

SpacemiT has now officially launched the K3 Pico-ITX SBC and K3-CoM260 system-on-module with the RVA23-compliant, SpacemiT K3 octa-core X100 CPU with up to 60 TOPS of AI performance, up to 32GB LPDDR5, 256GB UFS, and PCIe Gen3 x4 NVMe SSD support. The board also features an eDP connector, a 10GbE SFP+ cage, a Gigabit Ethernet RJ45 port, built-in WiFi 6 and Bluetooth 5.2 wireless connectivity, two USB Type-C connectors, four USB 2.0 ports, an M.2 Key-B socket coupled with a NanoSIM card slot for 4G LTE or 5G cellular connectivity, and more. K3 Pico-ITX SBC specifications: System-on-Module - K3-CoM260 SoC - SpacemiT K3 CPU 8x 64-bit RISC-V X100 "big" cores clocked up to 2.4 GHz, RVA23 compliance; 130 KDMIPS performance (similar to RK3588) 8x RISC-V A100 AI Cores with support for up to 1024-bit RVV1.0 parallel computing, optimized for matrix operations. GPU - Imagination Technologies BXM4-64-MC1 GPU with Vulkan 1.3, OpenCL

CNX Software - Embedded Systems News
cnx-software.com [Unofficial]Apr 9

Sipeed T256s – A USB thermal camera with 256×192 LWIR sensor, 640×480 AI super resolution

https://fed.brid.gy/r/https://www.cnx-software.com/2026/04/09/sipeed-t256s-a-usb-thermal-camera-with-256x192-lwir-sensor-up-to-640x480-super-ai-resolution/

Sipeed T256s – A USB thermal camera with 256×192 LWIR sensor, 640×480 AI super resolution

Sipeed T256s is a portable USB thermal camera equipped with a native 256×192 Long-Wave Infrared (LWIR) sensor and a built-in 2.4 TOPS NPU for hardware-level AI super-resolution (ISR) up to 640×480 resolution in real time, effectively suppressing image noise without external software. The UVC camera device supports standard output (Y16 raw and MJPEG) and features both male and female USB Type-C ports on opposite ends for connection to PCs and smartphones. It also includes a 1.69-inch touchscreen for standalone operation with live viewing, palette switching, hotspot tracking, and image capture. Other features include an optional macro lens (~5 cm focus) for inspecting small components, internal storage for snapshots, <50 mK thermal sensitivity, a 25 Hz frame rate, and a durable CNC aluminum enclosure for efficient heat dissipation. The device is designed for electronics R&D, industrial maintenance, and HVAC diagnostics. Sipeed T256s specifications: SoC - Axera AX620Q (updated, see comments section);

CNX Software - Embedded Systems News
YRabbitMar 30
The latest serializer—OVIDEO—is now up and running on #apicula!
It took some work to get it going—in this #fpga series, it only works with a differential output. #gowin #sipeed
YRabbitMar 2
Everyday implementation of DSP multipliers - checking the result on LEDs with what is on the calculator :) #fpga #sipeed#gowin#apicula
cnx-software.com [Unofficial]Feb 10

PicoClaw ultra-lightweight personal AI Assistant run on just 10MB of RAM

https://fed.brid.gy/r/https://www.cnx-software.com/2026/02/10/picoclaw-ultra-lightweight-personal-ai-assistant-run-on-just-10mb-of-ram/

PicoClaw ultra-lightweight personal AI Assistant run on just 10MB of RAM

PicoClaw is an ultra-lightweight personal AI Assistant designed to work on less than 10 MB RAM and suitable for resource-constrained embedded boards such as the Sipeed LicheeRV Nano SBC going for around $15 and powered by a SOPHGO SG2002 RISC-V SoC with 256MB on-chip DDR3. I keep reading news about the OpenClaw personal AI assistant, after first finding out about it when the Cubie A7S SBC was launched. OpenClaw (previously ClawdBot) clears your inbox, sends emails, manages your calendar, and checks you in for flights from WhatsApp, Telegram, or any chat app. It's been shown to run a range of hardware platforms, but it can be resource-intensive, and HKUDS created the nanobot ultra-lightweight personal AI assistant with about ~4,000 lines of Python code, or roughly 99% smaller than Clawdbot's 430k+ lines. PicoClaw further builds on the nanobot project, and has been "refactored from the ground up in Go through a

CNX Software - Embedded Systems News
Keldrim Feb 8

Decided to dig out my Lichee Pi 4A and update it's software (big pain in the ass) to see how much it has grown, way more snappier then I remember, which is still painfully slow but at least faster then the Starfive Vision 2. USB is a bit iffy but that was mentioned.

#Sipeed #RISC_V #RISCV

Keldrim Jan 29
And the cat is out of the bag #SpacemiT new K3 #RISC_V #RISCV CPU fully RVA23 compliant intruduced, did not spot a ITX board with a proper PCIe slot yet, from several announcements, but some things up for future grabs at #Sipeed
and the specs look very cool:
https://sipeed.com/k3
Sipeed

深圳矽速科技有限公司矽速科技致力于开源智能硬件、人工智能、边缘计算、图像处理等产品的研发、生产和销售，面向企业提供 AIoT 软硬件产品的行业整体解决方案，面向开发者提供开源软硬件平台，发布的一系列开源软硬件深受国内外开源社区开发者的广泛关注

cnx-software.com [Unofficial]Jan 27

MaixCAM2 modular 4K AI camera is based on Axera AX630 SoC with 3.2 TOPS NPU (Crowdfunding)

https://fed.brid.gy/r/https://www.cnx-software.com/2026/01/27/maixcam2-modular-4k-ai-camera-is-based-on-axera-ax630-soc-with-3-2-tops-npu/