Highlights:

  • The product family claims to deliver approximately double the performance per watt compared to the chipmaker’s previous generation of silicon.
  • The Xeon 6900P series includes four processors with 72 to 120 cores and higher base frequencies than the 6980P.

Recently, Intel Corp. has launched the Xeon 6900P series, a new series of server processors specifically designed for high-demand environments like artificial intelligence clusters.

The product line claims to deliver approximately double the performance per watt compared to the chipmaker’s previous generation of silicon. Intel reports that the Xeon 6900P series excels in handling AI workloads, with some chips in the lineup able to complete certain inference tasks 2.3 times faster than their predecessors.

Executive Vice President and General Manager of Intel’s Data Center and AI Group, Justin Hotard, said, “Demand for AI is leading to a massive transformation in the data center, and the industry is asking for choice in hardware, software and developer tools.”

Intel’s Xeon server chip portfolio, which includes the new Xeon 6900P series, utilizes two types of cores: one optimized for power efficiency and the other focused on performance. While some Xeon chip families feature a combination of performance- and efficiency-optimized cores, the new Xeon 6900P series is exclusively equipped with performance-oriented cores to enhance processing speeds.

The flagship processor in the series, the Xeon 6980P, boasts 128 cores that function at a base frequency of two gigahertz. For brief periods during intensive workloads, these cores can nearly double their speed to 3.9 gigahertz. Additionally, the cores are backed by a 504-megabyte L3 cache, which stores data actively used by the applications during operation.

The Xeon 6900P series also features four additional processors with lower computing capacity, offering 72 to 120 cores that are configured with higher base frequencies than the flagship 6980P.

The five chips in the series share a unified design. The cores of a 6900P processor are built on three distinct silicon pieces, or chiplets, which also house the chip’s cache and other related components. The cores can utilize the cache as a shared storage space or divide it to maintain their data in separate memory pools.

Intel manufactures the chiplets using its latest Intel 3 manufacturing process, which is the company’s second to utilize extreme ultraviolet lithography (EUV) technology. This process offers 18% improved performance per watt compared to Intel’s first-generation EUV implementation.

The three chiplets housing the cores and cache of a Xeon 6900P processor are integrated with two additional semiconductor modules manufactured using the earlier Intel 7 node. These modules enhance tasks like data compression and encryption, and they also include the input and output (I/O) circuits that enable the chip to connect with other components of the server in which it is installed.

Another advantage of the Xeon 6900P series is its support for MRDIMM memory, a faster variant of DDR5 that is commonly used in data centers. MRDIMM offers the potential for up to 39% more bandwidth compared to earlier memory technologies.

The technology also simplifies chip manufacturing in several ways. It features a so-called tall form factor (TFF) configuration, which effectively doubles the maximum memory capacity a processor can support. Additionally, it achieves this without the need for complex chip packaging components that would drive up manufacturing costs.

Along with the launch of the Xeon 6900 series, Intel today officially unveiled the Gaudi 3 machine learning accelerator, which it introduced in April. This chip is aimed at providing an alternative to Nvidia Corp.’s dominant graphics processing units. Intel claims that the Gaudi 3 can execute inference tasks 30% faster than Nvidia’s previous-generation H200 GPU.

Internally, Intel’s new AI chip consists of two groups of computing modules. It includes eight MME modules optimized for executing relatively simple machine learning tasks, as well as 64 TPC units designed to handle advanced AI workloads, such as large language models.

Intel has released a reference architecture for an AI appliance capable of housing up to 256 Gaudi 3 chips. The company states that organizations with advanced needs can connect multiple such appliances to form a single cluster. AI clusters built in this way can support up to 8,000 Gaudi 3 chips.