【🔥 Edge AI Highlights】
1. Community Expert Achieves: Raspberry Pi 5 Runs Llama 3 (8B) at 10 tokens/sec!
Hot posts on Hacker News show that developers have successfully run the 8 billion parameter Llama 3 model on Raspberry Pi 5 using community-optimized quantization frameworks (such as GGUF Q4_K), achieving a speed of 10 tokens/sec in practical tests.
【💬 Sharp Review】:
This is no longer a question of “can it run,” but rather “can it be used.” A speed of 10 t/s means that a smooth local AI assistant (like a smart home control center) on edge devices has become a reality. This is thanks to the community’s extreme optimization of quantization technology. The Raspberry Pi 5 (Cortex-A76) is becoming the hottest “toy” and validation platform for large models on the edge.
【🤖 Latest Developments in Large Models】
2. Mistral AI Secures $500 Million in New Funding, Announces Launch of “Mistral Medium” Edge-Optimized Model
French AI unicorn Mistral AI is unstoppable. Just after announcing its massive new funding, it quickly unveiled its “Mistral Medium” model. The company claims that this model has been deeply optimized to balance performance and power consumption, specifically designed for “on-device” inference, targeting the AI PC market of Qualcomm and Apple.
【⚙️ Chips and Hardware】
3. (Key Point) NXP Expands S32 Automotive Platform, Releases S32K39 MCU Focused on xEV Motor Control
NXP has released the latest member of its S32K3 series MCUs—the S32K39. This chip is designed for motor control and inverters in new energy vehicles (xEV), integrating a high-performance Cortex-M7 core, ASIL D functional safety features, and a hardware security engine (HSE) for secure OTA updates, making it a typical high-end dedicated MCU.
4. SiFive Partners with CEVA to Create High-Performance RISC-V Edge AI Platform
RISC-V leader SiFive has announced a partnership with DSP giant CEVA. The two will integrate SiFive’s RISC-V CPU IP with CEVA’s AI/CV DSP IP to provide a customizable, high-performance heterogeneous platform for automotive, industrial, and mobile edge AI applications.
5. STMicroelectronics Q3 Financial Report: STM32 (including U5 series) Demand Remains Strong
In last night’s earnings call, ST revealed that despite a weak consumer electronics market, orders for its industrial and automotive-grade STM32 MCUs (especially the ultra-low-power U5 series) remain robust, demonstrating the strong resilience of the high-end MCU market.
【💡 Innovations and Applications】
6. Zephyr 3.8 RC2 Released: Enhanced LoRaWAN and Matter Support
The Zephyr RTOS project team has released the second release candidate (RC2) of version 3.8. This version focuses on strengthening support for the LoRaWAN protocol stack and improving compatibility with the Matter 1.3 specification, clearing obstacles for building cross-ecosystem smart home and IoT devices.
7. New Features in AI Smart Locks: User Recognition via “Knock Pattern”
A new type of smart lock has emerged, featuring an AI coprocessor. In addition to fingerprint and NFC recognition, it can identify homeowners by analyzing the vibration waveforms and rhythms of knocking sounds (a kind of “vibration fingerprint”), addressing the issue of wet hands not being able to recognize fingerprints.
【🔭 Crowdfunding Trends】
8. “Mooltipass ReMM”: Open Source Hardware Password Manager
An open-source hardware password manager project called Mooltipass ReMM has gained popularity on Kickstarter. It uses an STM32L4 core, has passed FIDO2 certification, can store and manage passwords offline, and communicates with PCs via an encrypted USB protocol, aiming to provide physical-level digital security, having raised over 100,000 euros so far.
【🛠️ Daily Toolbox/Tips】
9. GitHub Hotspot: GDB-Dashboard
If you are still using the plain black frame GDB to debug embedded code, you must try this. It is a Python script (gdb-dashboard) that transforms your GDB session into a modular TUI (Terminal User Interface) dashboard, displaying registers, disassembly, memory, and stack information in real-time, greatly enhancing debugging efficiency.
【👇 Geek Question】
10. Today’s Discussion: The NXP S32K39 is designed for motor control. Do you think the future of high-end MCUs lies in “one core to rule them all” (like M85) or “dedicated is king” (like S32K39)?
A. Dedicated is king! Motor, safety, AI, each has its own role to play for reliability.
B. One core to rule them all! The computing power of M85/M55 is sufficient, dedicated chips are too costly.
C. Software-defined hardware, the future is heterogeneous, both are important.
D. (Feel free to leave your thoughts in the comments!)