Sapphire Rapids Is A Critical Turning Point For Intel

With the launch of the 4^th generation of Intel Xeon Scalable processors and the Xeon CPU Max series, which were codenamed Sapphire Rapids, Intel is charting a new server architecture path for the future that will have a significant impact on Intel and the rest of the industry. The launch includes 52 new product SKUs aimed at specific server workloads, including artificial intelligence (AI), networking, 5G Radio Area Networks (RANs), data encryption and security, and High-Performance Computing (HPC). Intel achieves this through a variety of CPU architectural enhancements, the integration of ten functional accelerators, new instruction extensions, and in some cases, the integration of High-Bandwidth Memory (HBM). The company even transitioned to using a chiplet (multi-die) configuration. The result is a series of products that looks and performs differently than any other Intel product before it.

The purpose of this article is to highlight the reason for Intel’s dramatic change in server processor design and the likely impact of the change. First up is the why. The x86 is a mature architecture that Intel has been reinventing and evolving for more than four decades. It is the most successful high-compute architecture ever produced partly due to its general-purpose nature and broad industry support. However, the nature of computing, especially in the data center, has been changing. Not only have the types of different workloads increased but many workloads have become more specific and defined, especially as scaling has resulted in thermal and power challenges. This, combined with challenges in lithography scaling, one of the factors influencing Moore’s Law, have resulted in smaller performance improvements from one generation to the next.

ADVERTISEMENT

Consequently, these changes lead to a different way to architect processors – to include dedicated hardware accelerators that perform specific functions better and more efficiently than the general-purpose CPU, much like the benefits attributed to external accelerators like GPUs. Ironically, this change moves away from the SPECint performance benchmarks Intel has touted for the past several decades. These new processors will not win any SPECint performance awards, but as many IT professional will indicate, SPECint is just the first step in evaluating the potential for a new processor. The ultimate performance test is running the intended workloads, which Intel is counting on for future design wins. According to Intel, the intended workloads will see a 50% to 1000% improvement in performance.

Intel has apparently been shipping some of these new Xeon products to select customers for more than a year and is just now offering general availability to the broader market. One reason has been manufacturing challenges with the products. Some of Intel’s manufacturing issues have been highly publicized. However, another issue is likely software. Adding accelerators incurs significant software overhead to ensure that the transition to the new architecture is seamless. This additional software includes support by the operating systems and applications that are critical to the intended applications. Please note that this has not been confirmed or denied by Intel. It is purely speculation by TIRIAS Research, but Intel also announced over 400 verified models as part of its AI Software Suite, which represents a significant investment in new software just for AI workloads.

ADVERTISEMENT

Another key change is the shift to on-demand upgrades to the processors. This will allow customers to enable certain features/accelerators at later date or only when needed. It’s still unclear how this will be received at the server purchasing level, but it fits right into the hyperscale service model where customers are paying for the time and the resources required for each workload.

The other key point is the impact of the new products. These products took two years longer to bring to the broader market than initially planned. However, I would call this Intel’s “Zen moment.” Several years ago, Intel’s primary competitor, AMD, went back to the drawing board to develop a new server architecture called Zen, which led to the now highly successful Epyc server processors. AMD made a bold move in incorporating architectural features from many other server processors, including those employed by Intel like multithreading. In addition, the company made a controversial move to using a multi-die chiplet architecture. Like any new silicon product, not every intended design function or feature made it into the first generation. In fact, the second generation resulted in major design changes. The third generation that is shipping now and is a highly refined and very competitive family of server processors. A positive byproduct Intel’s product delay delay has been the development of an extensive validation suite that is tried and tested for real world workloads.

The 4^th/Sapphire Rapids generation of Xeon Scalable and Xeon CPU Max processors represent a similar bold move by Intel. We are likely to see significant changes to the architecture over the next few generations that will further refine and enhance the performance of the product family. One might be to separate the accelerators from the CPU die to offer increased flexibility and customization. Another might be other types of accelerators for other workloads, including custom accelerators from customers or 3^rd parties. But no matter what the enhancements are to come, this is a major transition for Intel in many respects and one that will change the competitive dynamics of the market from one of just raw performance to one of workload performance. Tirias Research believes this is a change in the right direction because it corresponds to changes in server and datacenter architecture that will feature other processing/accelerator architectures and further customization to increase the performance efficiency. It may also impact the decision of customers to buy private cloud solutions vs. renting time on public cloud solutions for various workloads. However, adoption of a new processor architecture, especially in servers, usually takes several years.

ADVERTISEMENT