At present, the development of the digital economy has become a key force in changing the global competitive landscape. As computing power becomes the new engine of the digital economy, the scale of computing power continues to grow, and the structure of computing power has changed. Actively embracing the wave of intelligent computing, continuously outputting high-quality computing power to support the construction of digital China, and adapting ubiquitous and heterogeneous computing power to promote digital and intelligent transformation are becoming important considerations for new changes in future infrastructure.
Those who want millet work when they want to govern, and those who want to rule according to the situation. Facing new changes in future infrastructure, China Mobile grasps the trend of a new round of technological revolution and industrial transformation, focuses on "computing power + connectivity + capabilities", adheres to the concept of "openness + win-win", and is committed to creating a set of standard architectures and nurturing a Technology ecology, jointly create a series of original technologies to break through the current bottlenecks that hinder the healthy development of computing power infrastructure such as "closed technology ecology, complex technology stacks, and fragmented technology systems", and create efficient, open, controllable, and The computing power structure of the letter.
The mobile cloud COCA software and hardware integrated on-chip computing architecture is released
On April 25th, at the sub-forum of the 2023 Mobile Cloud Conference, China Mobile officially released the mobile cloud COCA (Compute on chip Architecture) integrated hardware and software on-chip computing architecture (hereinafter referred to as: mobile cloud COCA architecture), creating a national-level independent and controllable The high-performance computing power base drives the mature development of the localized intelligent computing industry from the three aspects of laying out the localized intelligent computing ecosystem, co-building an independent GPU+DPU industry alliance, and jointly creating high-performance network technology, and grasping the right to define cloud computing technology. Lead the cloud computing market in the next golden decade.
The mobile cloud COCA architecture is composed of three core units: GPU, DPU, and HPN. With COCA as the core, a large model based on technologies such as self-developed programmable DPU, multiple heterogeneous intelligent computing power, high-performance RDMA network, and Diskless storage architecture engine can be built. The computing power infrastructure aims to realize the horizontal integration and vertical abstraction of high-performance computing power clusters, and unify the acceleration of hardware offloading of computing, storage, network, security, and management and control capabilities.
- Heterogeneous collaboration of COCA-GPU units, building the cornerstone of intelligent computing infrastructure capabilities
In terms of GPU units, Mobile Cloud hopes to build an independent GPU joint ecosystem with partners, unify operator interface standards, provide a unified northbound interface, and be compatible with the deep learning framework of each GPU. Reduce customer migration costs through AI abstraction; improve training and inference efficiency through AI acceleration; improve GPU resource utilization through AI pooling.
2. The COCA-DPU unit realizes the integration of software and hardware, releasing the surging performance of intelligent computing infrastructure
In terms of the DPU unit, based on the COCA architecture, five offloading engines, including the network engine, storage engine, management and control engine, virtualization engine, and security engine, have been built. The integration of software and hardware has restructured the computing power infrastructure, helping the programming acceleration unit DPU to become a new computing power. The power core redefines the new standard of cloud computing technology in the era of computing power, and builds a new technology curve in the era of computing power.
3. The COCA-HPN unit overcomes the industry problem of multi-machine multi-card scale interconnection
In terms of HPN units, the COCA architecture integrates technologies such as X-Link + self-developed "Wumeng Protocol" + distributed communication library. "Protocol" focuses on creating end-to-end high-performance network connections, and the distributed communication library focuses on computing power scheduling and connections. The three cooperate to tackle the supply of computing power in application scenarios such as large-scale models, break through the bottleneck of computing power connection, and promote the capacity of new computing power infrastructure.
COCA software and hardware integrated on-chip computing architecture makes the "rock stronger"
The COCA architecture is mainly oriented to the technical ecology of computing power network infrastructure. Through computing power unloading, decoupling, and lossless technologies, it strives to build a standardized access ecosystem for computing power, and at the same time bridge the gap between computing power and applications.
At the 2023 Mobile Cloud Conference, China Mobile released the third-generation Panshi server, which realizes independent controllability of key technologies, integration of software and hardware, and is now fully commercially available to support computing-intensive scenarios. The Panrock DPU on the Panrock server is managed by the COCA DPU unit, which can effectively improve the core performance of the overall computing power, overall storage, and read and write resources.
At present, COCA has access to more than 30 partners, harvested more than 50 national-level independent intellectual property achievements, and supported more than 10 industrial implementation cases, covering government affairs, finance, scientific computing and other fields. In order to accelerate the prosperity and development of the COCA ecology, a COCA joint ecological working group has been established. In 2024, the three major units of DPU, GPU, and HPN will be opened up, and a full-stack autonomous and controllable system will be created for large-scale model scenarios to help the country's new high-tech performance infrastructure.
Facing the future, Mobile Cloud will continue to break through the problems of independent innovation of cloud infrastructure root technology, build a new type of large-scale computing power infrastructure based on COCA, and work with industry partners to promote the prosperity of the technology ecosystem, drive the mature development of the localized intelligent computing industry, and contribute to the construction of a digital China Contribute more power.