CXL Roadmap and Intel’s Plans
Cloud computing, data center infrastructure, and high-performance computing (HPC) segments are seeing a return to typical growth patterns as a result of the fervor surrounding Artificial Intelligence (AI)-related products like ChatGPT, a generative AI software.
With this trend, one of the most talked-about technologies is the quick uptake of Compute Express Link (CXL) products and solutions. A selection of the 2023 CXL announcements are as follows:
- CXL swap out for Xconn
- Solution for CXL test validation from Teledyne Lecroy
- MemVerge’s CXL collaboration software for multi-server infrastructures
- Intel Xeon scalable CPUs, fourth and fifth generation, supporting CXL 1.1
- Release of Revision 3.1 of the CXL standard
- Introduction of numerous additional goods that comply with CXL 1.1
Additionally, Intel FPGAs helped to propel CXL forward by announcing two firsts for the industry:
- The first FPGA to accomplish compliance with CXL 1.1 for Type 1, Type 2, and Type 3 devices.
- The first-ever real-world demonstration of FPGA acceleration through the use of a CXL Type 2 interface with the newly released 5th generation Intel Xeon scalable CPU.
A University of Illinois at Urbana-Champaign team implemented the new Intel CXL Type 2 demonstration. By offloading the Linux algorithm function kernel same-page merging (KSM) to an Intel Agilex 7 FPGA over a CXL Type 2 link to a 5th Gen Intel Xeon scalable CPU, a Redis database application was accelerated. Using the Yahoo! Cloud Serving Benchmark produced remarkable gains in performance:
- Running the KSM function and DB application on up to 76% less CPU cycles.
- Reduced tail latency by up to 90% for read/write operations in memory.
These are unmistakable performance and power reduction outcomes that provide system suppliers reduced total cost of ownership and enhanced sustainability, as well as quicker task turnaround times for end users. This is significant because, in order to meet demand, solution providers have been forced to integrate AI accelerators (GPU, CPU, ASIC/ASSP, and FPGA) into their entire system architecture due to the rapid uptake of AI applications. Investing in FPGA accelerators to offload compute-intensive workloads such as AI has continuously reduced total cost of ownership and power consumption. This benchmark attests to the much higher performance and throughput that CXL provides. As a result, we anticipate that in 2024 and 2025, many additional CXL-based milestones and solutions will be disclosed
You can develop your next-generation high-performance accelerators for a range of end applications using Intel’s CXL ingredients.
- Intel Agilex 7 I-series FPGAs (CXL 1.1 compliant)
- Intel Xeon Scalable Processors, Fourth and Fifth Generation (CXL 1.1 compliant), and associated CXL Intellectual Property (IP)
- A technical study on CXL written by the same group at the University of Illinois in Urbana-Champaign: Understanding CXL Memory with Real CXL-Ready Devices and Systems
