How generative AI is accelerating silicon diversity

As an IT leader, you probably haven’t spent a lot of time obsessing over the microprocessors powering your corporate computing hardware. Who can blame you?

It’s hard to master the intricacies of silicon semiconductors let alone allow them to rent space in your head when your priorities include modernizing your organization. A chef doesn’t stop to consider the electronic systems powering her culinary tools as she’s trying to plan and execute a multi-course menu.

Your organization likely uses laptops, PCs, and servers with Intel, AMD, and a few other lesser-known processor brands. What more do you need to know?

A bit more, actually. As your organization experiments with generative AI and other sophisticated computational jobs, what’s inside your server chassis and other computing machines is very important.

Seizing the opportunity, semiconductor companies have pledged to spend more than $200 billion on chip-related manufacturing projects in the U.S., according to the Semiconductor Industry Association.  

Silicon diversity is here to stay

These companies are accelerating silicon diversity, where servers and other computing appliances run multiple types of chips to power large language models (LLMs) that fuel generative AI tools, machine learning-based analytics, and high-performance computing systems that can ostensibly help you gain competitive advantages.

No one expects you to master the intricacies of nanometer manufacturing, but you should at least familiarize yourself with the different types of chips that power everything from smartphones and virtual reality devices to autonomous vehicles and sophisticated HPC clusters.

For instance, you likely know that central processing units (CPUs) are often used for general-purpose work, such as running operating systems, loading data, and managing memory. Tasks for which sequential processing is sufficient.

And while you may be aware that graphics processing units (GPUs) imbue gaming systems with their amazing look and feel, did you know that GPUs’ parallel processing capabilities make them essential for training and operating the LLMs that power everything from chat-based virtual assistants to image and video creation AI?

Crucially, CPUs and GPUs run in the same servers, offering a combination of memory and high performance required for some of today’s most demanding computational chores.

But wait—some servers contain even more chips! In addition to CPUs and GPUs, some machines feature data processing units (DPUs), which as their name implies are efficient at handling data-intensive workloads, such as data transfer, compression, and encryption.

DPUs can yield performance improvements for hefty workloads, including AI, ML, and HPC workloads while reducing power consumption due to their efficiency in processing data.

Enough, you’re thinking. No more chips, please. My digital transformation-captivated brain can’t take it.

Yet given the excitement over all things generative AI, it would be foolish to ignore neural processing units (NPUs), which are designed to boost ML and AI workloads by offloading those tasks from CPUs and GPUs. And as with DPUs, do so in a more energy-efficient manner.

NPUs, which can be included standalone in servers or embedded or CPUs or GPUs, use specialized hardware optimized for operations executed in neural networks – those brainy computer vision constructs.

What’s next for silicon? Chiplets.

So where is this chip bonanza headed? More of the same, plus some addition by subtraction.

That is to say, the emerging silicon model is moving toward so-called chiplet systems, an emerging approach to computer processing that breaks chips down into more modular components.

Some experts believe chiplets will eventually comprise hundreds or even thousands of CPUs, GPUs, DPUs, and NPUs in a way that reduces current yield and design limitations. Ideally, these chiplets will cost less to manufacture and boast greater flexibility and performance than current designs.

Chiplet systems are not yet ready for prime-time use—standardized interfaces and packaging challenges remain hurdles.

Picking partners—and tools

As you navigate this increasingly complex world of silicon diversity, you can only control what you can control. This includes the architecture choices you make as you lean into AI tools and other innovations to modernize and transform your organization. Our recently announced solutions, including Project Helix, span IT infrastructure, PCs, and professional services to help customers simplify and accelerate generative AI deployment. Here’s where you can learn more about Dell Generative AI Solutions.