Intel has revealed its new Xeon 6 processor at the Computex 2024 event in Taipei. The AI optimized CPU is designed to boost the performance of AI workloads in data centers. This new processor will come in two variants featuring E-core (efficient core) and P-core (performance core) microarchitectures. The Xeon 6 6700 E-core version is set for release today on June 4, with the higher-end 6900 P-core version debuting in Q3 2024 and additional iterations arriving in Q1 2025.
Intel CEO Pat Gelsinger explained these advancements during his keynote at Computex. He highlighted Intel's broad scope of innovation across the entire AI market spectrum, from semiconductor manufacturing to PCs, networks, Edge, and data centers.
Xeon 6 Specs and Capabilities
The Xeon 6 processor seeks to address the growing requirements of contemporary data centers, balancing performance and efficiency. It offers features such as increased core counts, improved memory bandwidth through DDR5, and advanced inter-socket bandwidth via UPI 2.0. Moreover, it includes support for Compute Express Link 2.0, a unified OS and firmware integration, among other enhancements.
The 6700 series features up to 144 Efficient-cores and 86 Performance-cores, with a TDP of 350W per CPU. It also supports 8-channel memory up to 6400 MT/s DDR5 and boasts 88 lanes of PCIe 5.0/CXL 2.0. On the other hand, the 6900 series includes up to 288 Efficient-cores and 128 Performance-cores, with a TDP reaching 500W per CPU. This variant allows for 12-channel memory up to 8800 MT/s MCR DIMM and has 96 lanes of PCIe 5.0/CXL 2.0.
Notably, the Xeon 6 E-core, known as Sierra Forest, is built on Intel's 3 process established on the Intel 7 node. The P-core counterpart, Granite Rapids, will launch next quarter. Both the E-cores and P-cores share a consistent architecture, utilizing a shared software stack and an open ecosystem of hardware and software vendors.
Release Strategy
To align with market demands, Intel will stagger the release of the Xeon 6 models. Ryan Tabrah, Vice President and General Manager for Intel's E-core product line, stated that this strategy aims to build market momentum and meet customer needs.
The first release, the Xeon 6 E-core, offers substantial improvements over previous generations, including enhanced rack-level consolidation (3-to-1), up to 4.2 times rack-level performance, and 2.6 times better performance per watt compared to the second-generation Xeon processors.
Launch of Gaudi 3 AI Accelerator
Aside from the Xeon 6, Intel also launched the Gaudi 3 AI accelerator. The Gaudi 3 accelerator, produced on a 5nm process, doubles the AI FP8 compute power and quadruples the AI compute for BF16, in addition to offering double the network bandwidth and a 1.5 times increase in memory bandwidth compared to Gaudi 2.
It delivers 1,835 teraflops of FP8 compute performance, 128GB of HBM2e, and 3.7TBps of HBM bandwidth through its OAM-compliant accelerator card. Additionally, the Universal Baseboard offers 14.6 petaflops of FP8 compute performance, more than 1TB of HBM2e, and 29.6 TBps of HBM bandwidth. The latest PCIe CEM option for Gaudi 3 includes 1,835 teraflops of FP8 compute performance, 128GB of HBM2e, and a TDP of 600W.
The Gaudi 3 AI kit, which includes eight Gaudi 3 accelerators and a Universal Baseboard, is priced at $125,000. In comparison, the Gaudi 2 kit costs $65,000, also bundling eight accelerators and a Universal Baseboard. This is significantly more affordable than comparable systems from competitors such as Nvidia, which can exceed $300,000.
