Joinwin Electronics HK Limited518031About Join WinJoinwin Electronics HK Limited FLAT/RM 29, 22/F, YAN’S TOWER, 25-27 WONG CHUK HANG ROAD,ABERDEEN HONGKONG Shenzhen ZhongYiYingTong Technology Limited Room 1610, Block A, Overseas Decorative Building, Zhenhua Road,Huaqiang North Street, Shenzhen, 518031Join-Win will be your one-stop purchasing assistant. Join together, achieve win-win!Joinwin Electronics HK Limited518031About Join WinJoinwin Electronics HK Limited FLAT/RM 29, 22/F, YAN’S TOWER, 25-27 WONG CHUK HANG ROAD,ABERDEEN HONGKONG Shenzhen ZhongYiYingTong Technology Limited Room 1610, Block A, Overseas Decorative Building, Zhenhua Road,Huaqiang North Street, Shenzhen, 518031Join-Win will be your one-stop purchasing assistant. Join together, achieve win-win!Joinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics

Home
News
Talk About Intel Enterprise AI Strategy, And The Internal Logic OF The Fight Against Nvidia

Talk About Intel Enterprise AI Strategy, And The Internal Logic OF The Fight Against Nvidia

April 19 , 2024 0

Not long ago, we roughly combed Intel's development strategy in the era of generative AI. The context of the Intel Vision event is that generative AI technologies are critical to the future of enterprise customers. But we are in the early days of enterprise AI: enterprises now need AI solutions that are easy to deploy, scalable, and reliable and secure.

Another implicit background is that although there are already corresponding solutions on the market (NV...) However, the cost is high for enterprise customers (greenfield investment in infrastructure) and there is no choice. Therefore, Intel believes that whether from the perspective of customer demand, or from the perspective of market status, now is a good time for Intel to promote enterprise AI strategy. If we want to summarize the enterprise AI strategy that Intel talked about at the event, we think there are roughly three directions:

(1) From the cloud to the edge and end side, generative AI must reach;
(2) In an open way, from hardware to software, build an ecosystem around open standards with partners, and build an end-to-end generative AI architecture that gives users choice;
(3) From the system level, to solve the future performance requirements of generative AI, so Intel not only released Xeon 6 and Gaudi 3 processors, but also announced IPU, AI NIC and other networking related product plans, at the same time, but also prepared to focus on providing system-level reference design.

It may also be added that software is very important, so from the perspective of software and upper-layer applications, while emphasizing openness, emphasizing providing enterprise customers with more complete, easy to deploy, but also ensuring security and reliability of the solution. Readers who have not read the previous article are advised to take a look at Intel's enterprise AI strategic framework before reading this article.

In fact, during the Intel Vision event, Intel China also held a special media symposium to talk about the generative AI strategy announced by Intel with "open" as the keynote, how it is landed and reflected in China, and talked about Intel's more thinking about building an open ecology.

"Openness is a natural evolution"

We have been saying that the AI market pattern of the data center now determines that Intel must embrace openness and cooperation if it wants to play this deck of cards and win. When Li Ying, vice president of Intel Corporation and general manager of Intel's software and advanced technology business group in China, talked about the issue of open and open source software, he said it was "not a matter of choice, but a natural evolution, a natural development process."

"Pat (Intel CEO) released the picture of open platform, enterprise AI at the Vision event. It's rare for us to see a new alliance with that many players." Here Li Ying should be talking about the early participants of Intel's Open Platform for Enterprise AI (Enterprise AI), "I counted at least 20." "The scope of open, open source is going to be huge."

Intel's role in this open platform, "the first is compatibility, that is, how we provide the overall architecture to solve the interconnection and compatibility problems between different players, is the common embodiment of the value between everyone." Li Ying said, "The other point is more important, not to say that any technology must be provided by Intel. More importantly, we need to push the framework of standards, how to ensure that more and more open players are together, and how to ensure the overall direction of the process." This is in Li Ying's view, and the previous traditional similar framework is very different.

In addition, we mentioned that OpenVINO now supports the Arm platform, oneAPI supports NVIDIA Gpus, and here is an example. As mentioned in the previous article, when doing cross-GPU, cross-system, cross-node, Intel intends to use open standard high-speed SerDes, as well as new Ethernet-based protocols, to build AI infrastructure.

In fact, in July last year, a number of cloud service providers, chip manufacturers, and system vendors jointly formed the UEC Super Ethernet Alliance, hoping to build a complete Ethernet-based communication stack architecture for high-performance networking, mainly to adapt AI and HPC. Last year we wrote about the differences between Ethernet and Infiniband, talking about the weakness of native Ethernet for HPC, and Ultra Ethernet solutions are looking to combine the user base of Ethernet with the flexibility to serve AI and HPC in the new era.

Intel is definitely a member and supporter of UEC. "After the establishment of the UEC alliance, the corresponding white paper and the first version of the document were also published." "Intel has contributed a lot to this," said Yu Zhang, chief technology officer of Intel's Network and Edge business unit in China and senior principal AI engineer at Intel. "This protocol is open, Intel can use it, friends can use it, eco-partners can use it." "The benefit of openness is that the end user has more choices, so the cost can be reduced."
Here is an additional piece of information that was not mentioned in the previous part of this article. The AI NIC that Intel is expected to launch in 2026 will have two forms in the future: one is the form of an independent NIC; The other is chiplet, "different AI accelerators can integrate AI NIC (chiplet) into the accelerator chip." I wonder if Intel will sell this chiplet externally...

However, one thing that can be confirmed is that some media asked whether the IPU to be launched by Intel in the second half of the year supports third-party processors, and Intel's answer is also positive. "The network products we offer are flexible and open to different types, types and manufacturers of processors." Zhang Yu said.

In fact, the AI NIC expects to have chiplet existence, which can more or less see the initial form of Intel Products' joint Foundry business to engage in open ecology - meaning that there is a great probability to see Intel's AI NIC and other manufacturers' processors or accelerators do integration. It is only natural that the IPU supports three-party processors. I have to sigh that Intel is really different from Intel 10 years ago.

AI technology must be "accessible.
Based on the "open" idea, from the chip to the OEM system, to the software middleware, to the overall architecture of the upper application, the specific construction method, we may have to wait for the subsequent Intel reference platform. However, there are already examples of AI utilization of existing data center infrastructure.

This is also what Intel's enterprise AI rack concept emphasizes: there are options, brownfield investment based on existing infrastructure, and cost considerations. The logic behind Intel's approach is understandable, both because customers want access to new technologies at a lower cost and because Intel is actually the established market leader in data centers or enterprise IT infrastructure.

Since Nvidia has been loudly calling for accelerated computing to sweep all walks of life over the years, it is known that general-purpose processor cpus still have a place. And CPU, of course, is Intel's world.

"At present, only about 10% of enterprises have really deployed generative AI to enhance their competitiveness, and the other 90% have not yet fully deployed generative AI," said Zhuang Binghan (vice president of Intel's marketing group, general manager of data center sales in China and general manager of carrier sales in China) in a media Q&A session. ... "Much of the current corporate profit structure is based on general-purpose data center applications -- especially now that CPU iterations are so rapid, each CPU generation brings new performance improvements and power reductions."

"So there's still a huge need for universal data centers to grow their existing businesses." "Since the second half of last year, the construction of general data centers has returned to its original pace," Zhuang said. At the same time, even in the new generation of intelligent data centers, most of the infrastructure deploying Gpus and other accelerators still uses Intel cpus. (From this perspective, GB200/GH200 is really a greenfield investment in infrastructure)

"There are also companies looking to see if they can use cpus to do large-scale reasoning... In many cases, such as large models (inference) below the scale of 13 billion parameters, the CPU can do it." "We're also seeing customers who are willing to try to do large-scale reasoning with cpus."

"For enterprises, large models are also in the initial stages at this stage and don't require large GPU clusters to be deployed immediately - which is a big challenge for both operations and development." In addition, Intel now also offers GPU and AI chips, Zhuang Binghan stressed that intelligent data centers are also important market opportunities for Intel.

Liang Yali (vice president of Intel's Marketing Group and general manager of Cloud and Industry Solutions in China) mentioned in her speech the case of cooperation between Intel processors and AI solutions in the Chinese market, which seems to be very representative. For example, Intel and Jinshan Cloud cooperation, "in its seventh generation of performance assurance cloud server X7 to introduce the fourth generation Xeon scalable processor for targeted optimization." "We tuned three popular large models, Stable Diffusion, Llama2 and ChatGLM2, and released optimized model images."

AMX acceleration in Xeon processor improves the Stable Diffusion reasoning performance by 4.96 times; "On the Llama2 and ChatGLM2 models, the optimized model inference performance has also achieved a significant improvement of 2.62 times and 2.52 times, respectively." Intel believes that the cooperation with Jinshan Cloud reflects: first, "convenient deployment", there is no need to deploy AI inference servers; Second, "cost optimization"; Third, such a general-purpose data center can be flexibly used for other loads in addition to AI, "agile switching".

Jingdong Cloud based on the fifth generation Xeon processor to do intelligent marketing, intelligent customer service; In the field of intelligent manufacturing, cooperate with TCL Huaxing Optoelectronics to improve product yield; Education industry, and East China Normal University cooperation to develop large model integrated machine, improve teaching efficiency... All are Intel and Chinese customers, and at this stage, AI goes deep into different industries, which also confirms the specific cases of the stage of enterprise AI that Zhuang Binghan said.