Courtesy: Lam Research
With all the attention being given to AI, it’s easy to overlook some of the core technologies enabling its capabilities. Sure, a lot more people have now heard about NPUs, GPUs and the businesses that make them, but what about the companies that enable these cutting-edge AI accelerators to be manufactured?
The Complexity of Modern Chipmaking
While most people may not realise it, chip manufacturing is incredibly challenging and requires the level of scientific breakthroughs that have powered humanity’s most advanced achievements. I mean, we’re talking about bending the laws of physics to build components that are a thousand times smaller than a grain of sand. Oh, and doing so millions of times over at incredibly high levels of quality and consistency. Plus, with the extra demands that GenAI workloads are putting on today’s latest chips, the challenges are getting even tougher.
That’s why companies providing the equipment and technologies that enable the manufacturing of these advanced chips play an essential role in driving the advanced AI capabilities we are all starting to experience.
Without their work to overcome technical challenges like the need for exascale computing, addressing the “memory wall” that can slow down AI accelerators, increasing power efficiency, and other issues that are necessary to maintain the Moore’s Law-like advances we’ve seen in these chips, the state of AI would not be where it is today. In particular, organisations like Lam Research, which build extremely complex, sophisticated machines that help process the raw silicon wafers that eventually become today’s most powerful semiconductor chips, play a big, though little-understood, part in big tech advancements like AI.
Building Next-Generation AI Chips Through Heterogeneous Integration
Lam Research makes a wide array of equipment that performs multiple tasks in the highly precise, extremely complex, and long (often 30 days or more) process of creating a modern chip. But in the era of AI accelerators, it turns out even the most sophisticated individual chip isn’t enough.
Instead, the latest GPUs and other advanced processors are being assembled through a process called heterogeneous integration, which combines multiple independent elements, known as “chiplets,” into even more sophisticated pseudo-SOCs, or Systems on Chip (advanced multi-chip packages that mimic some characteristics of an SOC). Commonly referred to as advanced packaging, the technology that enables the creation of these pseudo-SOCs requires extremely sophisticated semiconductor manufacturing.
Extraordinarily precise component stacking, chip-to-chip connections, and other key technologies allow these chips to integrate multiple independent processing elements, separate connectivity elements, memory, and more. The ultimate goal is to create the most powerful and capable multi-chip package they can in the most effective and efficient space and power envelopes possible.
Advanced Packaging Techniques
As with individual wafer processing, there are often multiple steps and multiple technologies (and approaches) involved with chip packaging. Some entail direct side-by-side connections between various chiplets and other elements, while others use various forms of stacking technology where different pieces sit on top of one another. In all cases, a critical part of the packaging process involves creating the paths through which the connections between the various elements are made. Sometimes those paths are created through film layers that act as a type of “glue” between the elements, while in other situations, it may involve creating millions of tiny holes that are filled with a metal-like material that provides something akin to a physical bridge between the layers.
In the case of Lam Research, the company has developed machines for each of those core packaging technologies. For physical bridging types—which are called through silicon vias or TSVs—Lam offers products in their Syndion, Striker ALD, and SABRE 3D lines. Each performs different parts of the process, including etching for creating the holes, deposition and filling for both lining and then injecting the new material into the holes, and then various cleaning processes along the way.
Semiconductor Manufacturing Innovations Enable AI Progress
Though little understood, the advancements in AI acceleration that have been achieved to date are strongly tied back to the manufacturing technologies that enabled them to be built. Integrating things like High Bandwidth Memory (HBM) directly beside GPU cores, for example, has had a huge impact on the performance, scale and efficiency of the latest AI accelerators, and that, in turn, is driving the impressive advancements we’ve seen in Large Language Models (LLMs) and other AI applications.
Looking forward, it’s going to be continued advancements in 3D packaging—along the lines of what Lam Research is doing with their new VECTOR TEOS 3D tool—that allow those advancements to continue. They may not be easy to see, understand, or appreciate, but semiconductor manufacturing technologies play an enormously important role in moving the tech industry and society forward.
