With all that’s being said about the growth in demand for AI, it’s no surprise that the topics of powering all that AI infrastructure and eking out every ounce of efficiency from these multi-million-dollar deployments are hot on the minds of those running the systems.  Each data center, be it a complete facility or a floor or room in a multi-use facility, has a power budget.  The question is how to get the most out of that power budget?

Key Challenges in Managing Power Consumption of AI Models

High Energy Demand: AI models, especially deep learning networks, require substantial computational power for training and inference, predominantly handled by GPUs. These GPUs consume large amounts of electricity, significantly increasing the overall energy demands on data centers. AI and machine learning workloads are reported to double computing power needs every six months​. The continuous operation of AI models, processing vast amounts of data around the clock, exacerbates this issue, increasing both operational costs and energy consumption​.  Remember, it’s not just model training, but also inferencing and model experimentation​ which consume power and computing resources.

Cooling Requirements: With great power comes great heat.  In addition to the total power demand increasing, the power density (i.e. kW/rack) is climbing rapidly, necessitating innovative and efficient cooling systems to maintain optimal operating temperatures. Cooling systems themselves consume a significant portion of the energy, with the International Energy Agency reporting that cooling consumed as much energy as the computing! Each function accounted for 40% of data center electricity demand with the remaining 20% from other equipment.​

Scalability and Efficiency: Scaling AI applications increases the need for more computational resources, memory, and data storage, leading to higher energy consumption. Efficiently scaling AI infrastructure while keeping energy use in check is complex​.  Processor performance has grown faster than memory and storage’s ability to feed the processors, leading to the “Memory Wall” as a barrier to deriving high utilization of the processors’ capabilities. Unless the memory wall can be broken, users are left with a sub-optimal deployment of many under-utilized, power-eating GPUs to do the work.

Balancing AI Innovation with Sustainability

Optimizing Data Management: Rapidly growing datasets that are surpassing the Petabyte scale equal rapidly growing opportunities to find efficiencies in handling the data.  Tried and true data reduction techniques such as deduplication and compression can significantly decrease computational load, storage footprint and energy usage – if they are performed efficiently. Technologies like SSDs with computational storage capabilities enhance data compression and accelerate processing, reducing overall energy consumption. Data preparation, through curation and pruning help in several ways – (1) reducing the data transferred across the networks, (2) reducing total data set sizes, (3) distributing part of the processing tasks and the heat that goes with them, and (4) reducing GPU cycles spent on data organization​.

Leveraging Energy-Efficient Hardware: Utilizing domain-specific compute resources instead of relying on the traditional general-purpose CPUs.  Domain-specific processors are optimized for a specific set of functions (such as storage, memory, or networking functions) and may utilize a combination of right-sized processor cores (as enabled by Arm with their portfolio of processor cores, known for their reduced power consumption and higher efficiency, which can be integrated into system-on-chip components), hardware state machines (such as compression/decompression engines), and specialty IP blocks. Even within GPUs, there are various classes of GPUs, each optimized for specific functions. Those optimized for AI tasks, such as NVIDIA’s A100 Tensor Core GPUs, enhance performance for AI/ML while maintaining energy efficiency.

Adopting Green Data Center Practices: Investing in energy-efficient data center infrastructure, such as advanced cooling systems and renewable energy sources, can mitigate the environmental impact. Data centers consume up to 50 times more energy per floor space than conventional office buildings, making efficiency improvements critical. Leveraging cloud-based solutions can enhance resource utilization and scalability, reducing the physical footprint and associated energy consumption of data centers​.

3. Innovative Solutions to Energy Consumption in AI Infrastructure

Computational Storage Drives: Computational storage solutions, such as those provided by ScaleFlux, integrate processing capabilities directly into the storage devices. This localization reduces the need for data to travel between storage and processing units, minimizing latency and energy consumption. By including right-sized, domain-specific processing engines in each drive, performance and capability scales linearly with each drive added to the system. Enhanced data processing capabilities on storage devices can accelerate tasks, reducing the time and energy required for computations​.

Distributed Computing: Distributed computing frameworks allow for the decentralization of computational tasks across multiple nodes or devices, optimizing resource utilization and reducing the burden on any single data center. This approach can balance workloads more effectively and reduce the overall energy consumption by leveraging multiple, possibly less energy-intensive, computational resources.

Expanded Memory via Compute Express Link (CXL): Compute Express Link (CXL) technology is specifically targeted at breaking the memory wall.  It enhances the efficiency of data processing by enabling faster communication between CPUs, GPUs, and memory. This expanded memory capability reduces latency and improves data access speeds, leading to more efficient processing and lower energy consumption. By optimizing the data pipeline between storage, memory, and computational units, CXL can significantly enhance performance while maintaining energy efficiency.

Liquid cooling and Immersion cooling: Liquid cooling and Immersion cooling (related, but not the same!) offer significant advantages over the fan-driven air cooling that the industry has grown up on.  Both offer means of cost-effectively and efficiently dissipating more heat and evening out temperatures in the latest power-dense GPU and HPC systems, where fans have run out of steam. 

In conclusion, balancing AI-driven innovation with sustainability requires a multifaceted approach, leveraging advanced technologies like computational storage drives, distributed computing, and expanded memory via CXL. These solutions can significantly reduce the energy consumption of AI infrastructure while maintaining high performance and operational efficiency. By addressing the challenges associated with power consumption and adopting innovative storage and processing technologies, data centers can achieve their sustainability goals and support the growing demands of AI and ML applications.

Explore AITechPark for top AI, IoT, Cybersecurity advancements, And amplify your reach through guest posts and link collaboration.

The post Balancing Brains and Brawn: AI Innovation Meets Sustainable Data Center Management first appeared on AI-Tech Park.

Whether you’re choosing a restaurant or deciding where to live, data lets you make better decisions in your everyday life. If you want to buy a new TV, for example, you might spend hours looking up ratings, reading expert reviews, scouring blogs and social media, researching the warranties and return policies of different stores and brands, and learning about different types of technologies. Ultimately, the decision you make is a reflection of the data you have. And if you don’t have the data—or if your data is bad—you probably won’t make the best possible choice.

In the workplace, a lack of quality data can lead to disastrous results. The darker side of AI is filled with bias, hallucinations, and untrustworthy results—often driven by poor-quality data.

The reality is that data fuels AI, so if we want to improve AI, we need to start with data. AI doesn’t have emotion. It takes whatever data you feed it and uses it to provide results. One recent Enterprise Strategy Group research report noted, “Data is food for AI, and what’s true for humans is also true for AI: You are what you eat. Or, in this case, the better the data, the better the AI.”

But AI doesn’t know if its models are fed good or bad data— which is why it’s crucial to focus on improving the data quality to get the best results from AI for IT use cases.

Quality is the leading challenge identified by business stakeholders

When asked about the obstacles their organization has faced while implementing AI, 31% of business stakeholders involved with AI infrastructure purchases had a clear #1 answer: the lack of quality data. In fact, data quality ranked as a higher concern than costs, data privacy, and other challenges.

Why does data quality matter so much? Consider OpenAI’s GPT 4, which scored in the 92nd percentile and above on three medical exams, which failed two of the three tests. GPT 4 is trained on larger and more recent datasets, which makes a substantial difference.

An AI fueled by poor-quality data isn’t accurate or trustworthy. Garbage in, garbage out, as the saying goes. And if you can’t trust your AI, how can you expect your IT team to use it to complement and simplify their efforts?

The many downsides of using poor-quality data to train IT-related AI models

As you dig deeper into the trust issue, it’s important to understand that many employees are inherently wary of AI, as with any new technology. In this case, however, the reluctance is often justified.

Anyone who spends five minutes playing around with a generative AI tool (and asking it to explain its answers) will likely see that hallucinations and bias in AI are commonplace. This is one reason why the top challenges of implementing AI include difficulty validating results and employee hesitancy to trust recommendations.

While price isn’t typically the primary concern regarding data, there is still a significant price cost to training and fine-tuning AI on poor-quality data. The computational resources needed for modern AI aren’t cheap, as any CIO will tell you. If you’re using valuable server time to crunch low-quality data, you’re wasting your budget on building an untrustworthy AI. So starting with well-structured data is imperative.

Four facets of high-quality, trustworthy data for IT use cases

To understand why the quality of data matters, let’s look at AI in IT—an area that has value for nearly every industry. New AI models for IT can reduce the number of help tickets, dramatically lower the time needed to resolve problems and help you make better decisions by proactively highlighting potential issues before purchasing new software. In a field where a mistake can cost your organization millions of dollars at scale, a good AI solution is worth its weight in gold. But how do you ensure that it’s using good data?

The first thing to consider is the breadth of data. More data across more sources typically makes an AI more trustworthy, as long as you’re collecting good data. Think of it this way: a single restaurant review can offer a glimpse into its quality, but a restaurant with numerous reviews provides a more accurate assessment, allowing you to make a more informed decision. Was the one negative issue an outlier? Or is there a pattern that should be identified and evaluated?  Similarly, an AI trained for IT on 10,000 data points collected every 15 seconds from endpoints will be more useful than an AI trained on 800 data points every 15 minutes.

Next, focus on data depth. The amount of data a model has from IT endpoints can make a significant difference. In one instance, a company had 3,000 systems crash after a software patch didn’t play nice within the existing setup. The IT team quickly resolved the issue using a patented AI that identifies correlations between their system changes and device anomalies. This process was possible because the AI had been trained on their unique datasets, including historical data.

As AI trained for IT collects data, it’s crucial that the data is well-structured and as clean as possible. Most data sets will invariably have some noise—data that’s meaningless, irrelevant, or (in some cases) even corrupt, but training AI on high-quality, well-structured label makes all the difference. 

Finally, don’t forget about your people. AI is simply a tool. Change management and the impact of AI on humans are invaluable considerations when making decisions about introducing AI capabilities and use cases and (perhaps most important) evaluating if the AI you’re using for IT is delivering the most useful results for your organization. As AI continues to transform nearly every industry, think of data as the ingredients in the best AI recipe. If the ingredients are bland, the power and nuance of AI is lost. AI that’s fed robust, rich data, however, delivers on all the promises and opportunities of well-trained models. From there, the results will follow.

Explore AITechPark for top AI, IoT, Cybersecurity advancements, And amplify your reach through guest posts and link collaboration.

The post How to improve AI for IT by focusing on data quality first appeared on AI-Tech Park.