The Future Of Next-Generation HPC At Scale: The Transformative Impact Of PCI Express 8.0

Abstract

High-Performance Computing and Artificial Intelligence are getting even closer, at unprecedented scales, building the computational models that exascale and the next generation zettascale architectures will require. PCI Express 8.0 is the next disruptive evolution of the de facto standard for high bandwidth, low latency interconnect that will help heterogeneous computing overcome the very limits that have historically restricted its ascendance. The specification also includes Pulse Amplitude Modulation four-level (PAM4) signaling, improved Forward Error Correction (FEC), and improved protocol stacks, which result in meaningful improvement of the effective data rate. Modern computers adopt more and more diversified accelerators such as Graphics Processing Unit (GPU), Field-Programmable Gate Array (FPGA), or domain-specific accelerated processors with complex interconnect topologies. PCIe 8.0's high bandwidth and low latency empower system architectures to efficiently keep nodes well fed with data to achieve the desired system performance. These system architectures may utilize cache-coherent memory fabrics, disaggregated pools of resources, and different peer-to-peer communication modes to power the next-generation systems enabled by PCIe 8.0. Furthermore, the implementation of such approaches at scale encounters difficulties related to power delivery network design, thermal management, and channel implementation in order to maintain the integrity of the transmitted information. Such implementations have been shown to provide varying benefits to applications such as streaming data processing and machine-learning workloads in distributed systems. In latency-sensitive scientific computing applications, the speed advantage is less impressive, and system integration considerations such as new cooling methods, power management, and signal conditioning become major success factors. As a vendor-neutral and open industry specification, PCIe 8.0 benefits from interoperability and multivendor support across the broad PCI Express ecosystem. The PCIe 8.0 specification will support the next generation of computing systems from the present generation of exascale systems to future systems based on zettascale technology.