news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

• Cressey from Aofei Temple
  Quantum Bit | Public Account QbitAI

Artificial intelligence is undoubtedly an important topic today, and it is driving the development of the digital economy through various means such as big models, vertical applications, and intelligent agents.

However, the infrastructure that serves AI——Computing power and data, which is the bottleneck restricting the development of artificial intelligence, has also become a topic of further discussion.

Why does this bottleneck still exist? And how can we break it?

We have seen some answers in what Nortel Digital Intelligence has done since its establishment a year ago.

An "AI factory" that integrates domestic computing power that is "ready to use right out of the box" has surfaced in Beijing.

“AI Factory” targets computing power anxiety

Whether it is for a single practitioner or the entire industry,Lack of effective computing powerThese are all pain points that need to be solved urgently.

From the demand side, the booming development of generative AI represented by large models, the fierce "Battle of Hundred Models", and the coming of the era of reasoning will undoubtedly bring about a huge increase in demand for computing power.

A report jointly published by the Chinese Academy of Engineering and more than a dozen other institutions in the Science journal Intelligent Computing pointed out that the computing power required for artificial intelligence doubles every 100 days.It will increase more than 1 million times in the next five years。

Moreover, in the era of artificial intelligence, the demand for computing power is no longer just the number of FLOPS; the structure of computing power demand has also changed significantly.

In the AIGC era, the structure of computing power has changed from CPU+accelerator to GPU-dominatedLarge-scale parallel computing, the proportion of intelligent computing power is increasing rapidly。

However, at present, domestic chips, especially high-end chips, have generational gaps with the industry's advanced products in many aspects, such as computing power performance, power consumption wall, and inter-card communication speed.

Non-chip factors are also limiting the efficient use of computing power, such as insufficient network transmission capabilities between nodes, an imperfect supporting software ecosystem, and so on.

Under this reality, most of the domestic intelligent computing power cannot be used efficiently, giving rise to the industry's second anxiety about computing power.

The root cause of this anxiety isUsers do not have a clear understanding of the capabilities of multi-core computing, especially domestic computing power, which can be summarized into several aspects:

• Domestic large-scale multi-core clusters lack the ability to support large-scale computing power adaptation and evaluation close to the production environment. In other words, single-chip testing can only provide estimated data for interconnection between cards and computing nodes, but there is no data on acceleration ratios;
• The existing platform mainly uses index-based evaluation indicators, which cannot be connected with user business scenarios and fail to fully open up the adaptation link from business scenarios to underlying chips. In other words, actual users cannot understand and judge whether the multi-core cluster can meet their needs.
• There is a lack of multi-core heterogeneous service platforms in China, as well as attempts at multi-core heterogeneous, hybrid training, and reasoning tasks.

In general, demanders do not know the services that suppliers can provide and the extent of support for their actual business. Suppliers lack a platform to demonstrate their capabilities and what they can do, as well as a fair and clear evaluation and standard.

In order to break this perception and solve the puzzle of computing power anxiety, a“Spark Smart Computing”The AIDC brand came into being.

“A single spark can start a prairie fire.” The solution provider chose the name “Spark” because they hope that the brand can start a prairie fire and be replicated across the country.

The one who proposed this plan wasNortel Digital Intelligence——A high-tech enterprise focusing on artificial intelligence under Beijing Electric Control.

Nortel Digital Intelligence believes thatComputing power is a new form of productivity in the era of artificial intelligence：

In the agricultural era, the land was computing power;
In the machine age, electricity and steam power are computing power;
In the era of artificial intelligence, computing power is the primary productive force, just like coal, water and electricity.

In order to turn computing power into a productive factor like electricity, we can no longer just focus on simply stacking chips, servers, and cabinets;Transforming from a “computing warehouse” to an “AI factory”, equipped with corresponding new quality production relations.

This is also the new idea of ​​Nortel Digital Intelligence in creating Spark Intelligent Computing.

Specifically, Spark Intelligent Computing supports the innovation and application of artificial intelligence technology by providing universal computing power, just like water, electricity and coal.

On the other hand, with the concept of "AIDC as a Service", we provide supporting services such as base big models, industry big models and one-stop tool chains, so that companies can easily and quickly build the desired applications or reconstruct current products to achieve innovation in production methods.

At present, the first benchmark project of the "Xinghuo Intelligent Computing" brand isBeijing Digital Economy Computing CenterConstruction has already started and the project will be promoted nationwide as a template in the future.

To better assist in computing power upgrades, Nortel Digital Intelligence has two core products that fully address the computing power anxiety in the industry:

• Advance AI Heterogeneous Computing Platform, in response to the current shortage of high-end computing resources, we configure a multi-core heterogeneous solution that best meets the needs of the domestic market;
• Pagoda Large Model Adaptation Platform, widely adapt the chip downward, so that the chip and the large model are widely decoupled and adapted;

The two products are "two sides of the same coin", solving computing power anxiety from different dimensions.

As a full-stack domestic computing power iteration verification platform, the Qianjin AI heterogeneous computing platform can achieve unified management and scheduling of multiple computing powers. At the same time, it is the first to promote effect-based billing. Users can apply on demand without having to understand the underlying details, and achieve smooth migration between multiple computing powers.

In order to make better use of computing power, Spark Intelligent Computing has also made breakthroughs in unified communications, compilation optimization and other technologies. The Advance AI heterogeneous computing platform canImprove domestic chip performance by 2-3 times and throughput by 6-7 times。

While providing computing power support for users, the Qianjin AI heterogeneous computing platform also createsA new evaluation mode of "chip × model × scenario", helping to break the information asymmetry between manufacturers and users, and helping chip manufacturers break through,Let domestic chips go from "usable" to "easy to use"。

This means that while providing computing power support to users, it is also promoting the comprehensive improvement of domestic chips and bringing more source computing power.

The Pagoda Big Model Adaptation System is an enabling model and application side, supports model deployment and management, performs refined operational management of computing power, and further supports artificial intelligence applications.

Let high-quality data flow

If the lack of computing power is a "dark cloud" hanging over the AI ​​building, then the lack of data is another well-deserved one.

Dou Dejing, chief scientist of Nortel Digital Intelligence and distinguished professor of the School of Computer Science and Technology at Fudan University, said in a public speech,Data scale and quality determine the "upper limit" of AI——Whether it is training or reasoning, high-quality data is needed as support.

Taking Meta's latest open source large model Llama 3.1 as an example, its third-generation model consumes 15T tokens of training data on more than 16,000 H100 GPUs, which is 7 times that of the second generation. This also achieved a performance leap between the two generations at the same parameter scale.

In my country, the current situation of data supply side is also facing the same anxiety as computing power.The total amount is abundant, but there are many high-quality data silos。

According to IDC estimates, China's data volume will grow from 23.88 ZB to 76.6 ZB between 2022 and 2027, ranking first in the world. However, available high-quality corpus data sets are still relatively scarce.

With the shortage of data, high-value data is facing“Can’t supply, can’t circulate, can’t be used well”The dilemma of data resources making it difficult to realize their value makes the already limited data resources even more tense.

On the other hand, the degree of data integration is seriously insufficient, with a large amount of data scattered across industries or vertical systems.Not structured, cannot be directly used for large model training.

The deep-seated reason behind this isLack of mature technology and trusted mechanisms to support data security, companies that own data "don't know, don't dare, can't, and won't" let the data flow freely. This "four no" situation has further created today's shortage of data sets and formed a vicious circle.

So the most urgent task isEstablish a reliable and secure data mechanism。

To this end, Nortel Digital Intelligence launched another important product:Red Lake Trusted Data Space, allowing data elements to be fully gathered and allowing the scenes to be fully open.

Honghu Trusted Data Space is a multi-level solution built around "trust", aiming to activate data to provide high-quality trusted data services and create a data environment of "secure sharing and trusted circulation".

From a technical perspective, the Red Lake Trusted Data Space is equipped with a trusted data sandbox, which uses technologies such as "logical data warehouse" to import, modify and delete data, and uses privacy computing + blockchain technology to ensure data security and leakage-free, as well as data traceability and transparency, to achieve "full-chain control".

This approach makes data credible, controllable, available, auditable, and encrypted, thereby creating a data sharing and application environment where multiple parties trust each other and data circulation agreements are fulfilled.

Only with such controllable security measures throughout the entire process can data owners dare to put their data resources into market circulation, and encourage more data owners to participate in it.

With the initial data investment, the data supply ecosystem will mature, the form of data sharing will become more complete and convenient, and the flowing high-quality data will inject more fresh blood into the development of the industry.

At the same time, Nortel Digital Intelligence also brings high-quality data to the industry.A vertical model matrix has been laid out in industries such as medical care, education, and culture, allowing artificial intelligence companies to better find application scenarios.

So, who exactly is Nortel Digital Intelligence, and how did it come up with such a comprehensive set of solutions that target the pain points of the artificial intelligence industry?

The "power plant" in the era of artificial intelligence, the "bead stringer" of the industrial chain

Nortel Digital Intelligence is a high-tech enterprise under Beijing Electric Control Group, focusing on artificial intelligence.Full-stack AI service provider。

Before Nortel Digital Intelligence was established, GPT-3.5 was released. At that time, a new round of artificial intelligence era was in full swing, and AI competition was becoming increasingly fierce. In response to the development of the times and market demand, Nortel Digital Intelligence came into being.

When it was founded, the first project undertaken by Nortel Digital Intelligence was the Beijing Digital Economy Computing Center, which was positioned asProviding intelligent computing power to the public and achieving universal access to computing power。

In other words, this computing center is like a "power plant", allowing computing power to empower the entire AI industry layering from base to application like water, electricity and coal.

Head of Digital Strategy and Market, NortelYang ZhenIntroduction: This is the difference between Nortel Digital Intelligence and Digital Technology Company. The first priority of Nortel Digital Intelligence is to support the artificial intelligence industry so that artificial intelligence companies or individual developers can quickly use and enjoy artificial intelligence.

In the entire artificial intelligence industry, Yang Zhen believes that Nortel Digital Intelligence plays aThe "bead stringer" in the industry chainThe role of the industry chain is to string beads together if there are already beads in the industry chain; if not, make them.

For example, the acceleration of domestic chips, the layout of vertical models in industry models, and the trusted data space mentioned above are all "pearls" in this industrial chain.

In addition to providing computing power products, the Beijing Digital Economy Computing Power Center has also specially designed communication and exhibition spaces to allow AI practitioners in the same or different fields to communicate and exchange ideas, and promote more comprehensive and in-depth innovation.

In the future, the open space in this computing center will also be open to the public as a "Computing Technology Park" to allow more people to understand the power of production transformation in the era of artificial intelligence.

In short, Nortel Digital Intelligence will use the intelligent computing center as a support to connect upstream and downstream technologies, products and services such as chips, servers, big models, AI applications, and industrial investment, promote the growth and prosperity of the artificial intelligence industry, and drive the upgrading of regional economic industries.

As of now, Nortel Digital Intelligence has more than a thousand AI industry ecosystem partners, covering star tracks such as AI-assisted drug development, general robots/embodied intelligence, unmanned driving, and AI intelligent bodies.