Intel XEON Scalable Processor Fourth Generation - Sapphire Rapids!

Intel has released details of the fourth generation XEON Scalable processors, the Sapphire Rapids family of Intel 7 server and workstation processors.

Five years in development, these new processors feature up to 60 cores, PCIe Gen 5, DDR5, HBM2e Memory, support for Compute Express Link 1.1, as well as a range of advanced features to accelerate computing including data streaming accelerator, QuickAssist Technology, and In-Memory Analytics Accelerator.

Overall, the new features in the Fourth Gen XEON processors are designed to be more energy efficient – delivering up to 2.9x more compute per watt. There are a range of features that optimise power and performance and deliver workload tuned performance for HPC/ AI / data analytics.

Watch this video on YouTube

Max Series For HPC and AI

For a range of XENON customers, there are some really interesting advances in the Sapphire Rapids based on workload requirements across HPC and AI. The XEON Max Series is the first x-86 based processor with high bandwidth memory, which will accelerate many native HPC workloads – with no coding required. With up to 64 GB of high bandwidth memory (HBM2e), there will be significantly more throughput on Sapphire Rapids clusters. Tests are showing up to 3.7x more performance in modelling and simulations.

The XEON Max series includes high core counts – 32, 40, 48, 52 and 56 core versions. These are all 2 socket designs pulling 350W.

Max Series GPUs – Ponte Vecchio

At the same time as the Sapphire rapids launch, Intel has announced their matching line of GPUs. Previously known by the code name Ponte Vecchio, these GPUs pack over 100 billion transistors, and can support up to 128GB of HBM2e memory, delivering up to 52 teraflops peak FP64 performance.

Max GPUs can be combined with Max CPUs in a 3:1 optimal ratio.

It will be interesting to see whether this tight design alignment of CPUs and GPUs will lead to optimised, accelerated platforms when these are available for test later in 2023.

On Demand (Software Defined Silicon)

Intel is introducing the option for field enabled capabilities to be activated with additional license fees. Called On Demand, or previously known as Software Defined Silicon, these On Demand features allow the Fourth Gen XEONs performance to be tailored and accelerated for specific workloads, and also to be adapted to future requirements with upgrades anytime in the future.

Intel On Demand applies to the following features: Dynamic Load Balancer, Data Streaming Accelerator, In-Memory Analytics Accelerator, Quick Assist Technology and Software Guard Extensions. These On Demand options are not available on the Max series CPUs, where the HMB2e and accelerators are integral to the Max CPU performance. These On Demand features will be available in the mainstream XEON Fourth Gen CPUs, where traditional corporate data centre workloads live.

Customer benefits from On Demand includes being able to access a more powerful CPU at a lower cost, and not having to pay for things that you may not need or want. It will be interesting to see how this plays out in the market, as it is a new approach and a new business model in this core segment of the server market where these CPUs will land. See HPCwire’s take on it here.

Cooling and Liquid Immersion

In previous blogs, we’ve highlighted how much power the new generation of CPUs and GPUs draw, and how much heat the generate. Conventional rack space design is now dominated by power and cooling concerns, often leading to less than fully populated racks. This decreasing density is more expensive and environmentally costly.

Immersion cooling has a lot of promise, but comes with a lot of its own problems.

With the launch, Intel also announced an immersion cooling warranty rider for its chips (with approved liquids); an open-IP reference design for immersion cooling; and partner-developed, rack-level solutions for cold-plate designs that support Sapphire Rapids. This includes OCP Specification for immersion liquids, validated CoolIT, HeeHee single Phase and Zutacore 2 phase cooling. This follows last years announcement of Intel’s Open IP design for immersion cooling, and their investment of $700 million into immersion cooling solutions.

The company also highlighted its efforts to support software that intersects with workload efficiencies (e.g. Kubernetes Power Manager, Kubernetes Telemetry Aware Scheduling, Intel Granulate).