dcbox小金库钱包·(中国区)官网

Recently, ONNX Runtime, the open-source community behind the renowned AI inference framework, released the official version 1.17.0, which includes support for LoongArch. Users can now use this release directly from the ONNX Runtime open-source community to develop and deploy AI inference applications on the Loongson platform, thereby enhancing the LoongArch AI ecosystem. ONNX Runtime (ORT) has gained significant popularity as an AI inference framework software in recent years. It serves as a foundational AI inference engine for numerous AI applications, supporting different model formats such as PyTorch, Tensorflow, and TFLite. In addition, it accommodates various computing backends, including CPU, GPU, IoT, NPU, and FPGA. During the development of ORT version 1.17.0, Loongson's technology team closely collaborated with the ONNX Runtime community. They contributed 7697 lines of code to the community's code repository and conducted comprehensive vector optimization on core operators like matrix multiplication, convolution, and transposition. With the community's support, the LoongArch-optimized code successfully passed quality assurance processes, including inspection and test verification. As of ORT version 1.17.0, the ONNX Runtime community officially provides native support for LoongArch.

In September 2023, Loongson Technology completed the preliminary development of the printer main control chip—Loongson 2P0500. LS2P0500 is a main control SoC chip designed for the control system of single/multi-function printers and is the core controlling component in printing and scanning devices. This chip is used for print data reception, parsing, and processing; print engine control; timing control scanning; data scanning; image processing; motor control, etc. The chip employs heterogeneous CPU topology, incorporating one LA364 core and two LA132 cores, along with a 512KB shared L2 cache. It integrates various functional modules (such as DDR3, GMAC, OTG, USB, print interfaces, scanning interfaces, image units, PMIO, AD/DA, eMMC, SDIO, SPI, and PWM) and a power management and control module. Thus, the chip has wide application in printing, scanning, copying, and other typical scenarioses. Test results showed that all functions of the LS2P0500 are normal, with performance, power consumption, and reliability meeting expectations and requirements. Loongson Technology collaborates with multiple domestic mainstream printer manufacturers to meet the needs of various application scenarioses such as printing, scanning, and copying. Some prototypes are undergoing system testing by related manufacturers and are planned for trial production in small batches in the near future. Loongson Technology will complete the productization of the LS2P0500 by the end of 2023. It will keep delving into the field of printer chips and enhancing its cost performance and market competitiveness through series production and other complementary measures. Prototype Images

Recently, the new generation quad-core LoongArch-based CPU, LS3A6000, was successfully developed, which marked China's new milestone in the design of homegrown desktop CPUs. According to the test results from the electromagnetic compatibility laboratory of China Electronics Standardization Institute (CESI-EMC), the quad-core CPU, operating at a frequency of 2.5GHz, demonstrated a single-thread integer/floating point score of 43.1/54.6 in SPEC CPU 2006 base, and a multi-thread integer/floating point score of 155/140 in SPEC CPU 2006 base. The dual DDR4-3200 memory channels exhibited a bandwidth exceeding 42GB/s, with Unixbench measurements surpassing 7,400 points. These test results, evalsuated holistically, suggest that the overall performance of the LS3A6000 processor rivals that of Intel's 10th generation quad-core processors launched in 2020. Note: The test platform used for these results comprised a LS3A6000 with four cores and eight threads operating at 2.5GHz, 8GB DDR4 3200 memory, 256 NVME hard disks, and the Loongnix (V20.4) operating system alongside the LoongArch64 gcc8.3 compiler. The LS3A6000 processor is based on Loongson's proprietary instruction set architecture (LoongArchTM). This architecture, from high-level structure to instruction functionalities and ABI standards, was created totally with homegrown design and without licensing from any other country. Numerous international open-source communities associated with ISAs, as well as operating systems like UnionTech OS, Kylin OS, EulerOS, Anolis OS, and HarmonyOS, have lent their support to LoongArch. This backing, combined with foundational applications such as WPS, WeChat, QQ, DingTalk, and Tencent Meeting, has formed a basic software ecosystem parallel to X86 and ARM. Representing the inaugural product of Loongson's fourth-generation microarchitecture, the LS3A6000 integrates four high-performance 6-issue 64-bit LA664 processor cores. Its primary frequency reaches 2.5GHz, and it supports both 128-bit vector processing extension instructions (LSX) and 256-bit advanced vector processing extension instructions (LASX). The processor also features simultaneous multi-threading technology (SMT2), allowing for a total of eight logical cores. The LS3A6000 includes a DDR4-3200 dual-channel controller and a trusted security module, offering secure startup solutions and national cryptographic (SM2, SM3, SM4, etc.) application support. Compared to its predecessor—LS3A5000 desktop CPU—the LS3A6000 boosts single-thread performance by over 60% under the same process. The overall multi-thread performance has seen a significant growth, delivering a smoother user experience. Future Loongson server CPUs are expected to multiply the performance of the previous generation of 16-core LS3C5000 and 32-core LS3D5000 server CPUs. The LS3A6000, along with other LoongArch CPUs, also achieve remarkable level of software compatibility. Loongson recently submitted a patch to the Linux upstream community to support the detection and scheduling of hyper-threading features of the 3A6000 and enhance its features. The LS3A6000 further refines the level of software and hardware cooperation for binary translation, enabling the operation of more cross-platform applications. It caters to a wide range of large and complex desktop application scenarioses.