Sponsored Feature The datacenter sector has made great strides in addressing the challenge of sustainability over the past 10 years. Despite the enormous rise in global demand for digital services, the emissions generated by the much larger volumes of data now being created and stored have been prevented from running rampant.

That's thanks in no small part to the energy efficiency measures taken by datacenter operators and their ongoing migration to renewable power sources.

But the battle is far from won. Substantial challenges lie ahead if the industry is to abide by pledges to achieve net-zero status within the next 20 years. The Climate Neutral Datacenter Pact is pushing for the sector's electricity consumption to be derived 75 percent from renewable or carbon-free sources by the end of 2025 for example, achieving 100 percent five years after that.

Nor are carbon emissions the only facet of the challenge. Datacenter operators have also been battling with unstable energy pricing and a consequent impact on their profitability. Set against that is a conflicting need to build more energy-consuming processing muscle into their hosting infrastructure to accommodate the latest generation of power-hungry applications. This is especially the case for facilities that host a heavy proportion of High Performance Computing (HPC), Artificial Intelligence (AI) workloads.

Enterprises are deploying AI more than ever before - adoption is reckoned to have doubled since 2017 - while use of the technology has moved on from simple predictive tasks to compute-intensive generative processes.

LLMs need huge hike in compute power

The uncomfortable truth that the whole datacenter ecosystem now has to confront - from facility operators to the application developers that want to harness the technology - is that today's generative AI tools rely on enormous language models to understand and respond to complex challenges. This is all part of giving enterprises and consumers a richer and more accurate experience. The inevitable result is a huge hike in the level of computing power needed, and with it a massive increase in the amount of energy consumed and carbon emitted. With compute requirements only heading in one direction, there are obvious environmental implications unless something is done.

"When you look at the applications that demand most power, HPC has historically been near the top of the list," explains Stephan Gillich, Director of Artificial Intelligence and Technical Computing - GTM for Intel EMEA Datacenter Group. "HPC has an inevitable bearing on total cost of ownership (TCO) when planning a datacenter. Within TCO, energy consumption is very significant. Now with the rise of AI and ML, we have the same sort of challenge we had with HPC. Both have one thing in common - they are compute-intensive. The more you compute, the more energy you need. The more energy you use means more of a need to cool the equipment, which demands yet more energy. It's a multi-dimensional problem."

The pressure, Gillich says, is on both datacenter operators, and the DevOps and the system architect professionals that serve them, to balance efficiency with achieving desired outcomes: "Everyone must question whether the architecture is providing the right sort of performance for the required applications," he believes. "When it comes to AI applications, we are typically talking about a high level of performance. People need to be asking themselves how efficient they are in terms of their operations, not just in the performance of applications."

The Leibniz Supercomputing Centre (LRZ) near Munich as an example of best practice here with its holistic focus on energy consumption: "They have learned how to make their centers more efficient, for example with warm water cooling and energy reusage for heating and other purposes, like running chillers," Gillich notes.

Why architecture matters

Given the high levels of compute and storage capacity needed to process applications like AI, ML and HPC, and the knock-on effect on energy consumption, it is clearly incumbent on datacenter operators to seek out optimal power efficiency within their server architecture, and for the DevOps community to be working with the most power-efficient platforms available. But there isn't one size of solution that fits all circumstances. Any stance on energy needs to flex according to the task in hand.

"If you take AI as a workload, then the question is how to run that workload and develop for that workload as efficiently as possible," explains Gillich. "When you look at how people are using workloads in datacenters, there are three basic swim lanes. In one, the operator is hosting a lot of applications with AI as just a part of that. Then there are datacenters that are focused mainly on AI and HPC. Thirdly there are centers that just focus on AI, say running a large language model for training and inferencing. In each instance the architecture requirements will differ."

Intel's approach is all about bringing the appropriate solution to a particular problem, offering the right sort of accelerator technology depending on the datacenter's priority. The new range of 4th Generation Intel Xeon Scalable Processors was built with this aim in mind, helping datacenter owner operators to drive down energy bills and meet green targets, all the while giving system architects an environmentally friendly platform to work off.

The CPUs offer built-in acceleration and software optimisation as a way to deliver better performance compared to just growing the CPU core count. They've also been engineered to provide considerable improvement in performance per watt efficiency compared to previous generations of Intel Xeon chips. After all, more efficient CPU utilisation means lower electricity consumption and a better chance of hitting sustainability goals.

Below are some of the most important performance-enhancing and energy-saving innovations that the 4th Generation Intel Xeon Scalable Processors introduce.

Step forward Intel AMX

Intel AMX is a built-in accelerator that improves the performance of deep-learning training and inference on the CPU, making it well suited for workloads like natural-language processing and image recognition. This higher performance is achieved without a substantial rise in energy expended, meaning more can be done for less compared to previous iterations of Xeon CPUs.

AMX can be utilized to support AI straight out of the box, while further optimisation at the software level can deliver additional performance gains to help datacenter operators create the energy-saving efficiency they need.,

When compared with prior generations of Intel chips, systems built on 4th Gen Intel® Xeon® Scalable processors also deliver 2.9X improvements in average performance per watt efficiency for targeted workloads as measured by Intel benchmarks for example (see E1, E6 here for full performance metrics, results may vary), and up to 70-watt power savings per CPU in optimised power mode with minimal performance loss.

The new Optimised Power Mode can deliver up to 20 percent socket power savings with a less than 5 percent performance impact for selected workloads (see E6 here, results may vary). Innovations in air and liquid cooling reduce total datacenter energy consumption further, says Intel.

The manufacturing process for 4th Gen Xeon Processors uses 90 percent or more renewable electricity at sites with state-of-the-art water reclamation facilities.

"Our new CPUs are the foundation for AI workloads, particularly where it comes to inferencing and deep learning," states Gillich. "You can run those tasks on our CPUs very efficiently in the datacenter. We have implemented technology in the fourth generation of our CPU which is accelerating those workloads. It runs the operations that are necessary for AI in a very efficient way, giving you more performance for less energy. The new AMX extensions are particularly important for AI, and specifically for deep learning where you manipulate a lot of matrices."

Sustainability will undoubtedly be a key topic for the datacenter industry as long as it remains the backbone of today's digital world. And innovative technology solutions like 4th Generation Intel Xeon Scalable Processors are certainly needed if we are ever to solve the broader climate challenge.

The future is not all doom and gloom. There may be serious environmental implications from our increased reliance on AI, but there are always better ways to do things, as well as enormous opportunities for AI to be part of the solution. Real-time data collection combined with AI, for example, has been shown to help businesses identify areas for operational improvement to help reduce carbon emissions.

With further advances throughout the technology stack, right down to the level of the CPU, and with AI deployed in new and smarter ways, the opportunity for keep those sort of sustainability improvements ticking over is there for all to see.

Sponsored by Intel.

INTEL DISCLAIMER: Performance varies by use, configuration and other factors. Learn more on the Intel Performance Index site. Performance results are based on testing as of dates shown in configurations and may not reflect all publicly available updates. See backup for configuration details. No product or component can be absolutely secure. Your costs and results may vary. Intel technologies may require enabled hardware, software or service activation.

https://www.theregister.com//2023/11/22/meeting_the_global_need_for/