This Time It Is Different

The future depends on data-center computers that we cannot build today. We are all under pressure to find a way.

Investment experts cite the “This time it’s different” fallacy as a warning against ignoring historical patterns and the power of human psychology when analyzing market behavior. We might have reached a technological “This time it’s different” moment as we consider the next steps in AI computing.

Nvidia is shipping its latest-generation Rubin multi-chip GPUs for cutting-edge AI compute performance and efficiency, and the company has already published its roadmap to subsequent Rubin Ultra and 3D-stacking Feynman devices. What we don’t know is how to handle the thermal issues and data throughput challenges at the packaging and interconnect levels to access the silicon’s performance in the real world.

Historically, engineers treated interconnect design as secondary to the challenges presented by chip lithography. Creating the processor presented the hurdle; connecting it came relatively easily. But this time it’s different.

The Rubin GPU, which contains 336 billion transistors, already stretches the limits of today’s connectivity technologies. Its multi-chip module (MCM) comprises two compute dies and two I/O dies. Nvidia refers to the PCB laminate specification “M9” as an internal performance standard for high-bandwidth AI interconnects that moves beyond the maximum performance of most manufacturers’ standard product ranges. M9 specifies Df in the order of 0.0007 at 10GHz to ensure suitable signal integrity and minimize thermal problems. It dictates specific combinations of resin, copper and glass fiber that only a very small number of materials suppliers can fulfill.

Rubin Ultra servers will demand PCBs with Df in the 0.00035 range, calling for another upgrade in substrate materials performance. As the industry works out how to respond, higher-grade Q-glass can deliver low Df and Dk at high frequency, but its rigidity creates manufacturability challenges. New hyper very-low-profile (HVLP) copper foils also promise performance beyond today’s typical maximum limits.

PTFE could offer another way forward, but engineers must contend with processing difficulties, CTE mismatch and pending legislation restricting the use of PFAS chemicals. Current resin systems will likely adopt evolved blends that meet Rubin Ultra requirements, but this technology is approaching its performance limits.

The intended Feynman architecture proposes co-packaged optical interconnects through scale-up and scale-out interconnects, improving efficiency and eliminating external electro-optical connector losses. Interestingly, I have an old board on my desk from before the dotcom crunch of 2000 that contains embedded optical interconnects, although obviously built with technology more than two decades older.

Optics can bypass some board requirements, but pushing current copper foil grades and resin chemistries to their limits restricts the number of materials suppliers capable of delivering in volume. It also introduces manufacturability challenges that only a small number of board producers can handle.

That shrinking supplier pool creates two problems. It makes resilient supply chains harder to establish as protection against geopolitical uncertainty, and, more importantly, it creates a serious capacity constraint. AI server demand already compromises supply and manufacturing throughout the global PCB business and fundamentally reshapes the market. An even more rarefied supplier base would likely struggle to sustain the pace of investment required for new AI data centers.

As a species, humans have a strong record of overcoming technological obstacles that once appeared insurmountable. When we saw the end of Moore’s Law and the potential threat it posed to continued economic development, we engineered our way into the More than Moore era. When transistor geometries approached their minimum limits, we created the FinFET as a three-dimensional solution.

But this time it is different because we are pushing against the physical limits of materials, and no obvious solution sits on the horizon.

Scientific research almost guarantees that somebody, somewhere, is working on something that could deliver the breakthrough we need. Our built-in human insecurities make us sensitive to the possibility that we have reached the end of the road, despite our track record as problem solvers.

Today’s research communities connect more closely than at any time in the past, thanks to the communication technologies and channels now available. Sharing knowledge and experimental results may create the conditions for a breakthrough by focusing multiple minds on the same problem.

On the other hand, history provides plenty of examples of parallel invention, including calculus, the telephone and the computer mouse, while also reminding us that sometimes a singular mind sees patterns invisible to everyone else. Charles Darwin had the same instruments available to him as anyone else in his field. The way he thought about what he saw produced the breakthrough.

As we look for a way to overcome the connectivity challenges in our path right now, I’ll take the liberty of adapting another phrase to express the faith we must have in our ability to eventually produce a solution: “Cometh the hour, cometh the tech.”

It is said that technological progress leads us to know more and more about less and less. Over time, we gain increasingly specialized insights into the technologies we develop. We need those insights, but our future success will also depend on a coordinated, cross-disciplinary approach that brings together chipmaking, multi-chip packaging and PCB design to achieve the data throughput and signal integrity needed to realize Rubin Ultra and Feynman servers.

And, of course, because we are human, the even faster and more powerful generations that inevitably come next.End of article content 

Alun Morgan is technology ambassador at Ventec International Group (venteclaminates.com); alun.morgan@ventec-europe.com. His column runs monthly.