Intel Advanced Packaging for Bigger AI Chips

This week at the IEEE Electronic Components and Packaging Technology Conference, Intel unveiled that it is developing new chip packaging technology that will allow for bigger processors for AI.

With Moore’s Law slowing down, makers of advanced GPUs and other data center chips are having to add more silicon area to their products to keep up with the relentless rise of AI’s computing needs. But the maximum size of a single silicon chip is fixed at around 800 square millimeters (with one exception), so they’ve had to turn to advanced packaging technologies that integrate multiple pieces of silicon in a way that lets them act like a single chip.

Three of the innovations Intel unveiled at ECTC were aimed at tackling limitations in just how much silicon you can squeeze into a single package and how big that package can be. They include improvements to the technology Intel uses to link adjacent silicon dies together, a more accurate method for bonding silicon to the package substrate, and system to expand the size of a critical part of the package that remove heat. Together, the technologies enable the integration of more than 10,000 square millimeters of silicon within a package that can be bigger than 21,000 mm²—a massive area about the size of four and a half credit cards.

EMIB gets a 3D upgrade

One of the limitations on how much silicon can fit in a single package has to do with connecting a large number of silicon dies at their edges. Using an organic polymer package substrate to interconnect the silicon dies is the most affordable option, but a silicon substrate allows you to make more dense connections at these edges.

Intel’s solution, introduced more than five years ago, is to embed a small sliver of silicon in the organic package beneath the adjoining edges of the silicon dies. That sliver of silicon, called EMIB, is etched with fine interconnects that increase the density of connections beyond what the organic substrate can handle.

At ECTC, Intel unveiled the latest twist on the EMIB technology, called EMIB-T. In addition to the usual fine horizontal interconnects, EMIB-T provides relatively thick vertical copper connections called through-silicon vias, or TSVs. The TSVs allow power from the circuit-board below to directly connect to the chips above instead of having to route around the EMIB, reducing power lost by a longer journey. Additionally, EMIB-T contains a copper grid that acts as a ground plane to reduce noise in the power delivered due to process cores and other circuits suddenly ramping up their workloads.

“It sounds simple, but this is a technology that brings a lot of capability to us,” says Rahul Manepalli, vice president of substrate packaging technology at Intel. With it and the other technologies Intel described, a customer could connect silicon equivalent to more than 12 full size silicon dies—10,000 square millimeters of silicon—in a single package using 38 or more EMIB-T bridges.

Thermal control

Another technology Intel reported at ECTC that helps increase the size of packages is low-thermal-gradient thermal compression bonding. It’s a variant of the technology used today to attach silicon dies to organic substrates. Micrometer-scale bumps of solder are positioned on the substrate where they will connect to a silicon die. The die is then heated and pressed onto the microbumps, melting them and connecting the package’s interconnects to the silicon’s.

Because the silicon and the substrate expand at different rates when heated, engineers have to limit the inter-bump distance, or pitch. Additionally, the expansion difference makes it difficult to reliably make very large substrates full of lots of silicon dies, which is the direction AI processors need to go.

The new Intel tech makes the thermal expansion mismatch more predictable and manageable, says Manepalli. The result is that very-large substrates can be populated with dies. Alternatively, the same technology can be used to increase the density of connections to EMIB down to about one every 25 micrometers.

A flatter heat spreader

These bigger silicon assemblages will generate even more heat than today’s systems. So it’s critical that the heat’s pathway out of the silicon isn’t obstructed. An integrated piece of metal called a heat spreader is key to that, but making one big enough for these large packages is difficult. The package substrate can warp and the metal heat spreader itself might not stay perfectly flat; so it might not touch the tops of the hot dies it’s supposed to be sucking the heat from. Intel’s solution was to assemble the integrated heat spreader in parts instead of as one piece. This allowed it to add extra stiffening components among other things to keep everything in flat and in place.

“Keeping it flat at higher temperatures is a big benefit for reliability and yield,” says Manepalli.

Intel says the technologies are still in the in R&D stage and would not comment on when these technologies would debut commercially. However, they will likely have to arrive in the next few years for the Intel Foundry to compete with TSMC’s planned packaging expansion.