Пятнадцать байтов на стек от конца

By integrating SiMa.ai’s energy-efficient Modalix AI platform with Cisco’s new, robust and ruggedized IE3500 portfolio of switches, customers can now deploy powerful, production-grade edge AI solutions across manufacturing, logistics, and industrial automation use cases.

Key Benefits of the Solution

The SiMa.ai Modalix solution with Cisco’s IE3500 switches delivers several critical advantages for industrial customers:

• Real-Time AI Processing: The Modalix platform enables sophisticated AI models to run directly at the edge, eliminating latency issues associated with cloud-based processing and enabling real-time decision-making for critical industrial applications.
• Industrial-Grade Reliability: Cisco’s IE3500 switches are purpose-built for harsh industrial environments, providing the robust networking foundation necessary to support AI workloads in manufacturing facilities, oil and gas operations, and other demanding industrial settings.
• Simplified Deployment: IE3500 switches include power over ethernet (PoE) capability that eliminates the need for separate power supplies for cameras and SiMa.ai Modalix platforms reducing the number of connections and simplifying deployment.
• Secure management at the edge: Through Cisco Secure Equipment Access (SEA), administrators can securely access the Modalix platform as well as cameras and remotely manage them.
• Enhanced Operational Efficiency: The combination will enable advanced applications such as predictive maintenance, quality inspection, anomaly detection, and process optimization, helping organizations reduce downtime, improve product quality, and optimize resource utilization.

(отсюда)

https://www.automotiveworld.com/articles/software-centric-chips-anticipate-the-ai-defined-vehicle/

"SiMa.ai is positioning its unique MLSoC platform as the gateway for unlocking next-generation ADAS/AV performance. Importantly, it has been conceived specifically for automotive applications."

"SiMa.ai’s MLSoC has an ML performance of 50 TOPS and a ResNet-50 (an image classifying architecture) performance of more than 300 FPS per watt. The company claims latency can be reduced by a factor of ten, with all compute in Level 2/3 systems achieved at less than 25W, or less than 100W at Level 4."

"SiMa.ai designed the chip according to three foundation pillars: to be compatible with any computer vision and generative AI application using ML at the edge, regardless of vehicle type; to offer best-in-class performance per watt; and to provide an interface usable by anyone instead of a small demographic of hardware experts."

"In December 2024, SiMa.ai combined its AI/ML capabilities with silicon design and verification firm Synopsys to maximise customisation for automotive-centric IP, subsystems, chiplets, and SoCs."

Смотрите какую прелестную композицию можно соорудить из тысячи крохотных сопротивлений.

Прошу любить и жаловать: операционка FP/M для крошечных процессоров. Проект наконец дозрел до минимально законченного состояния. Есть файловая система во флеш-памяти и на SD-карточке, командная строка с возможностью редактирования, и главное - динамический загрузчик. То есть можно выполнять бинарники *.exe из файловой системы во флеше. Вот пример.



Исходники лежат на Гитхабе: github.com/fp-m/fpm-embedded

В разделе Release v0.1 можно скачать готовый бинарник fpm-rp2040-2mb.uf2 и сразу прошить его в плату с чипом RP2040 - если у вас имеется такая. После чего подключиться к консольному порту через USB (на скорости 115200) и играться с системой. Подробности смотрите на странице Getting Started.

Пока всё это работает только на процессоре RP2040. Список поддерживаемых плат смотрите на странице Supported Platforms.

Всё это доступно по лицензии MIT. Собрать из исходных текстов можно по инструкции: Build From Sources.

Проект открытый: присоединяйтесь, господа. Это только начало. Идею можно доразвить в самые разные интересные стороны.

Рассматриваем потроха новейшей российской ракеты Х-59м2а, 2024-го года выпуска. Вскрытие показало процессор 1879ВМ5Я NeuroMatrix и FPGA Altera Max10.

МЦСТ выложил на публику исходные тексты binutils для процессоров Эльбрус: binutils-2.39.tgz

Я переложил на Гитхаб для сохранности: binutils-elbrus

Заслуживающие внимания файлы, описывающие архитектуру Эльбрус:

• cpu-e2k.c
• elfxx-e2k.c
• elfxx-e2k-pm.c
• elf32-e2k.c
• elf32-e2k-pm.c
• elf64-e2k.c
• elf64-e2k-kpda.c
• elf64-e2k-pm.c
• eir.c
• elfcpp/e2k.h
• gas/config/tc-e2k.h
• gas/config/tc-e2k.c
• gas/config/tc-stub.h
• gas/config/tc-stub.c
• gold/e2k.cc
• include/elf/e2k.h
• include/opcode/e2k.h
• include/opcode/e2k/aau-regs.def
• include/opcode/e2k/mu-regs.def
• include/opcode/e2k/state-regs.def
• include/opcode/e2k/state-regs.def.new
• ld/emultempl/e2kelf.em
• opcodes/e2k-dis.c
• opcodes/e2k-opc.c
• opcodes/e2k-opc.h

Присутствуют забавные вещи, к примеру:

extern int hack_for_e2k;

EE Times:

“Customers are saying that having just the IP doesn’t help them, they need an environment where they can explore how the IP will behave in the final environment, completely up front, so they can work with the semiconductor partner and say, I need this IP, I need this many versions of it, these configurations of it,” De Schutter said. “They also want to start the software development well before the car comes out. Customers said the only true solution is being able to do all this pre-silicon.”

“We’ve shown the world how powerful our [accelerator] is, how power efficient it is, which are the key ingredients for automotive,” he said. “Bringing this together is great for us, of course, it’s exciting because it opens up new ways to go into the market.”

“It’s not just the accelerator part, getting a high performing, super power-efficient ML SoC together means you need to have a lot of other pieces connected to that,” he said. “Theoretically a company could say, give me your IP and I’ll build it into my product, but then how do they test that? How do they virtualise it? How do they make a test run on a chip that doesn’t exist yet?”

“That will help us define our path in more detail,” Kroeger said. “The automotive world needs exactly the product we’re trying to enable here, and this is the first step we are taking.”

Про нас пишут:

• Synopsys and SiMa.ai Announce Strategic Collaboration to Accelerate Development of Automotive Edge AI Solutions
• Synopsys, SiMa.ai team up to speed development of AI chips for cars

"Synopsys, among leading manufacturers of chip-design software, will collaborate with SiMa.ai, a company that has designed energy-efficient hardware and software capable of handling diverse AI functions used in cars."

"Synopsys users will gain access to SiMa.ai's intellectual property and Synopsys' tools for simulating how a chip and software will interact."

"The integrated Synopsys and SiMa.ai solution will be designed to enable:

• Early architecture exploration: Capability to guide automotive design engineers to optimally-choose performance, power, and software application requirements for custom or 3rd party SoC development.
• Shift-left software development: The solution will be optimized for multi-modal functions and GenAI, as well as to provide an end-to-end workflow including virtual prototyping and emulation to accelerate software/hardware integration and speed time to market.
• Cost-effective and differentiated in-vehicle experiences: Silicon-proven building blocks and subsystems for high-performance automotive ML SoCs and chiplets that are fully customizable to support diverse workloads, integrating SiMa.ai's leading, easy-to-use ML software and tools.
• Continuous upgradeability of automotive edge AI solutions: Applications with AI capabilities can be developed to support both current and future requirements. In addition, virtual prototypes can be used to develop and test over-the-air updates."

Однако работы привалило на мою голову. 😀

В комментах зашёл разговор про Raspberry Pi 5 по сравнению с Pi 4. Я решил сравнить потребление. Меряю в двух режимах:

• на холостом ходу
• при 100% загрузке всех процессоров

Загрузку процессоров создаю, включая режим Stress в утилите s-tui. Не подключены: видеоадаптер, Wi-Fi, устройства USB. Сравнение не совсем корректное, так как на Pi 5 стоит Линукс, а на Pi 4 - FreeBSD.

Raspberry Pi 5 под Ubuntu 24.10:



Raspberry Pi 4 под FreeBSD 14.0:



Перемножаем напряжения и токи, получаем итоговую табличку:

                  Idle      Max
Raspberry Pi 5    3.70 W    8.77 W
Raspberry Pi 4    2.12 W    5.05 W