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
Analysis Xeon 6: Why the performance core and energy Efficiency Core? How Do You Do That With 288 Cores?

Analysis Xeon 6: Why the performance core and energy Efficiency Core? How Do You Do That With 288 Cores?

June 11 , 2024 0

Although the electronic engineering album has never in-depth analyzed the architecture of Xeon processors before, with the interpretation of the core microarchitecture of the core/Core Ultra processor and its advanced packaging, it is more or less possible to see the reality of Xeon.
In fact, in the past few years in the era of Xeon Ice Lake and Sapphire Rapids, Intel server CPU products are quite passive, especially in the pursuit of high core number of application scenarios. Therefore, the report card of Intel's data center business in the past few years is indeed not good enough - even if Intel is still an absolute share of the server CPU market, the pressure comes from both AMD and Arm camps.
When Intel announced Xeon's roadmap, it was widely believed that Sierra Forest and Granite Rapids using the Intel 3 manufacturing process would be a new turning point. Not only has the manufacturing process caught up with the mainstream, but the separate design of P-core (performance core) and E-core (energy efficiency core) has identified different market needs, and the performance and energy efficiency are expected to achieve leaps.
Since the beginning of the Intel Innovation event last year, Intel has announced a lot of information about this generation of Xeon processors, including the announcement of the Xeon 6 processor at the Intel Vision event in April this year. Recently, Xeon 6 processor products were officially released, and more details have surfaced.

Xeon 6 Is Divided Into Two Versions

Intel has already predicted that the new Xeon 6 will have two different series using only E-core and only P-core - the former code-named Sierra Forest, the latter code-named Granite Rapids, to meet different market needs. Many foreign media believe that the existence of Sierra Forest is mainly to compete with the Arm camp - typical products such as Ampere Computing and Amazon's self-developed Graviton processors, reflected in its large number of cores.

We believe that Xeon 6's product planning has about this component, but it is not so simple. The presentation by Liang Yali, Vice President of Intel's Marketing Group and General Manager of Industry Solutions and Data Center Sales in China, summed up the target application scenarios for these two Xeon processors with different cores:

"Distributed general-purpose computing, represented by cloud-native microservices, is handled by E-core processors; "Heavy computing tasks such as AI and high-performance computing will be handled by P-core."

There is a higher-level logic to this: a single architecture cannot solve all of a customer's problems, or perfectly cover all business scenarios. At present, Moore's law is lagging, and it is already an industry consensus to do chip design based on application scenarios. Therefore, a more "targeted" architecture "for different business loads" is very in line with the trend of The Times.

And it's not just the different architectures of the CPU, but also accelerators like Gpus, ASics, FPgas. In our opinion, the two core versions of Xeon 6 are more like the concrete embodiment of Intel's XPU strategy in products such as cpus. This line of thinking is also in line with the current mainstream trend of "green" and "energy-efficient" computing.

Chen Baoli (vice president and general manager of Intel's Data Center and Artificial Intelligence Group in China) said that the growth in demand he has seen from data center customers in the past few years is AI and microservices. AI is characterized by high requirements for core performance, memory bandwidth and latency.

Microservices on the cloud are not so sensitive to single-core performance, but have higher requirements for "volume of services" - more sensitive to core density - which can also be seen from the AWS application example for Graviton. High core count or core density is important for microservices, container workloads.

So Intel designed two product lines for Xeon 6 to meet the performance and power efficiency requirements of different workloads, and the two versions share the underlying platform and software stack. The specific load requirements of P-core and E-core can be seen in the following figure:

Even though the E-core is called the so-called "energy efficiency core," the overall performance of the processor, in terms of web and microservices, networking, media codecs, data services and other loads, compared to the second-generation Xeon also has about three times the performance per watt.

Intel this time also finally generously gave Xeon 6 compared with the next door Epyc in some dimensions of performance comparison: P-core Xeon 6 is 3.7 times stronger in AI reasoning performance, E-core Xeon 6 is 1.3 times stronger in media transcoding performance per watt. Although these two numbers are not comprehensive, we believe that this does reverse the unfavorable situation of Xeon processors in the past two generations.

First Time With E-Core: Is Performance OK?

Although in modern edge and embedded markets, we have seen Intel processors that use E-core entirely, such as Snow Ridge and Parker Ridge. But this is the first time E-core has been used on a server CPU. The model released this time is Xeon 6700E - based on the model suffix we know that this is a CPU with an E-core core. There will be new models to follow, including the Xeon 6900P (based on the P-core) to be released next quarter, and more models for the 2025 Q1:6900E, 6700P/6500P/6 SoC/6300P.

From a digital model point of view, the 6700 and 6900 are two large series, or two platforms from an Intel design chip point of view. The 6700 with E-core has a maximum of 144 cores, and the P-core version has a maximum of 86 cores. The 6900 series uses the E-core up to 288 cores, and the P-core version 128 cores.

Looking at the number of cores, it is true that Xeon processors are instantly mainstream, especially on the Sierra Forest side; Even if we look at Xeon using P-core alone, 86 cores is already 30% more cores than the fifth generation Xeon - not to mention that there will be 128 cores in the future.

The following figure shows the configuration information about socket socket support, TDP, memory channel, PCIe/CXL, and UPI interconnection.

The 6700 series has a TDP of 350W for a single CPU, and the 6900 series has a TDP of 500W. It is worth mentioning that the Xiton 6 platform has MCR technology support, which can further improve the memory frequency - the 6700 series supports the memory frequency to 8000MT/s, and the 6900 series can be raised to 8800MT/s - the total memory bandwidth is 2.36 times that of the previous generation. In addition, the 6700P series supports 1, 2, 4, and 8 channel expansions in terms of slots.

Of course, the product level still needs to be specific to the SKU model: as mentioned earlier, this release is only the Xeon 6700E series. The SKUs of different products are shown below, covering the 6710E with 64 cores and the 6780E with 144 cores at the highest end. This series of processor products TDP power consumption from 205W-330W. Different core releases are not only differentiated by the binning process, which is discussed in detail in the later section of the Packaging architecture.

Systems And Ecological Support: Standardization And Industrialization

To sum up the above content, there are roughly two aspects. First, Intel Xeon 6 is divided into two lines, performance core and energy efficiency core, based on "adapting to diverse computing power" and taking into account universal requirements. Second, this new release of Xeon 6 with E-core achieves higher core density, better performance, energy efficiency and TCO, and matches the current concept of sustainable, green data centers.

We are very excited to see how this Xeon processor generation compares to the performance and efficiency of our competitors Epyc and Neoverse. Let's see how Intel's turnaround is going.

But part of Intel's advantage in the data center space is ecology. From the perspective of system ecology, we have previously written about Intel's open ecosystem strategy, which is bottom-up, from hardware to software openness: a major value for enterprise customers is significantly better costs. This part of this article will not be repeated.

On the day of the press conference, different roles in the upstream and downstream chain of the data center, including Jinshan Cloud, Inspur, NANDA General, memory technology, Dell, Lenovo, etc., were launched for the XDX 6 release station. International companies announcing the deployment of Xeon 6 or participating in early testing mentioned Xeon 6 performance and energy efficiency improvements such as ebay, SAP, Ericsson, Cineca...

For example, Jinshan Cloud ninth generation efficient cloud server SE9 is the first domestic cloud server released based on Xeon 6; Ntu general database solution based on Xeon to achieve higher performance and efficiency; And ODM manufacturer memory technology has also specially launched a "low-carbon energy-saving server" similar to Green PC throughout the entire life cycle.

Finally, it is worth mentioning that the press conference interview "liquid cooling" and more efficient system design attention is still very high, after all, with the expansion of chip scale and data center energy consumption more and more become a global burden, green data centers, PUE <1.25 and so on are contemporary hot topics.

Regarding the heat dissipation scheme and system design, Chen Baoli revealed that when the official announcement of the 6900 performance nuclear platform in Q3 this year, "we will also announce the supporting technical support launched with hardware partners." This also highlights the current era is very popular chip manufacturers to do the next system design trend, itself as part of the ecology.

For example, Intel has previously introduced reference design schemes in terms of cold plate and submerged liquid cooling through cooperation. Zhao Shuai, general manager of Inspur information server product line, said in an interview that the price of cold plate this year is 70 to 80 percent lower than three years ago; Therefore, it is an effective solution to work with Intel to make standards - "push into national standards" from group standards, and eventually become international standards, as well as the formation of industrialization.

Scale is the key to driving new designs. These are also becoming an important part of the development of data centers in addition to chips. It is also a key component of the "system" that Intel has repeatedly emphasized in the past year.