Physics Limits Interposer Line Lengths

Electrical interposers provide a convenient surface for mounting multiple chips within a single package, but even though interposer lines theoretically can be routed anywhere, insertion losses limit their practical length.

Lines on interposers — and on silicon interposers in particular — can be exceedingly narrow. Having a small cross-section makes such lines resistive, degrading signals the farther they travel. Designers may want to use long interposer lines, but the results are likely to be disappointing. “You can’t do that on an interposer,” said Kevin Donnelly, vice president of strategic marketing at Eliyan. “You can’t have long advanced-packaging traces.”

There are other challenges. Maintaining acceptable signal integrity requires that high-speed signals get help from ground lines and planes. But ground planes must be perforated by holes created through the manufacturing process, which complicates return-signal paths.

An optical interposer, in contrast, faces fewer of the limits of electrical interposers because optical signals can travel longer distances across larger interposers than are feasible for electrical. Optical interposers are intended for photonic communication within the data center, with the interposer handling all the photonics and the electrical circuits placed atop it. But they’re not going to replace electrical interposers anytime soon.

Reducing losses
Today there are two fundamental types of interposers in production — organic (also called RDL, or redistribution layer, because the technology initially was employed for redistributing signals in fan-out packaging) and silicon (inorganic). Organic interposers are much less expensive to produce, but the feature dimensions are larger than what can be achieved with a silicon interposer.

The use of silicon itself doesn’t mandate the thin lines. That’s due to the need to minimize the number of redistribution layers. “You could have wider signals, but then you would need a lot more signaling layers to get the same number of signals out,” explained Donnelly. “And people don’t want 8 or 10 layers. They want to make four-layer interposers, so they’re using skinny traces that are close together, which are resistive and capacitive and very lossy.”

While line dimensions are large by die standards, they’re small by PCB standards. “The minimum solution is a 2µm line width and 2µm line spacing for organic materials,” said Pax Wang, director for advanced packages at UMC. “For oxide-based dielectrics, it’s common to achieve 0.5µm for the line width and space.”

An organic interposer with 3µm metal thickness provides a line cross section of 6µm². For the distances signals would need to travel, this makes for a resistive line. “They’re resistant because their cross section is really tiny,” said Mike Kelly, vice president of chiplets and flip-chip packaging development and integration at Amkor Technology. “It’s a pretty thin pipe.”

The resistance degrades signal voltage over longer distances. “The insertion loss is usually what our customers are most worried about, which means they’re losing voltage,” said Kelly. Whether that length is a limitation depends on how the line will be used. Most lines simply cross the gap between two adjacent chips, which are about 75 to 100µm apart, but such lines could run as long as 7mm.

“On a chip, when you’re running digital traces around, you can drop in repeaters and buffers,” Donnelly explained. “You can’t do that on the analog signals that are running here. So a 5mm or 4mm long trace is a pretty long run for a simple bare piece of copper wire.”

Connections to HBM typically are longer due to the placement of the HBM signals. “Generally, the I/O bank for HBM is in the middle of the part,” said Kelly. “The traces tend to be longer, but the speeds are slower. HBM4 is going to start out at 6.4 Gbps per physical line. UCIe is specified all the way up to 32 Gbps, but it’s always a much shorter trace.”

The HBM stacks must be kept close to the processor to keep the lines short, even though the heat from that processor will adversely affect the memory. “Most customers would love to have their DRAM be a little further away if they could,” he said.

Even so, longer lines still may be employed. “Designers are looking for metal lines long enough to cross at least half the reticle,” said Wang.

Meanwhile, maskless lithography can make even narrower lines. “We can go down to 30nm,” said Ken MacWilliams, president of Multibeam. No one is making lines that thin for mainstream interposers today, but that kind of precision can be particularly helpful for correcting alignment issues with chiplets and bridges.

Transmission lines unlikely
Resistance is an important line characteristic, but it’s not the only consideration. Resistance is nominally insensitive to signal frequencies, but overall signal integrity is heavily dependent on how fast a signal switches. Signal integrity is a function of the overall impedance of the line as well as its length. Lines longer than a half wavelength generally must be considered as transmission lines, requiring controlled line impedance.

For example, 15mm is half the wavelength of a 10 GHz signal. That would be a very fast signal on an unusually long line, so for the most part, it’s not necessary to create controlled-impedance lines using microstrip or stripline techniques. “You can always make a transmission line, but it requires a more expensive interposer to have a reasonable amount of loss,” said Eliyan’s Donnelly.

RF circuits may require them, however. “We have seen customers that need RF that use [such techniques] for impedance control,” noted Bassel Haddad, senior vice president and general manager of SkyWater Technology.

Maintaining signal integrity
Signal integrity relies on a good solid ground, which generally is provided by ground planes. Such a plane can be its own layer in an interposer, where it serves three functions. “It serves for power, impedance control, and as a return path,” explained Haddad.

Controlled impedances largely maintain the quality of a signal with respect to its own reflections. But even shorter lines can experience interference or crosstalk from adjacent lines, and very fast interfaces can create a very noisy environment. So even though strict transmission lines may not be necessary, they still prove useful in protecting sensitive signals by surrounding them with ground.

“If it’s a long signal trace, designers will go to great lengths to manage the losses and keep ground metal surrounding that high-speed trace, whether that’s planes above and below or inter-digitated co-planar ground traces in the signal routing plane,” said Kelly.

But ground planes on interposers and chips are different from those on PCBs, which can have solid planes. “You’re talking about sort of a waffle grid with, let’s say, 50% metal,” said Kelly. “It’s hard to tolerate a contiguous ground plane because there are usually gases that need to be evacuated during the fabrication process. So you need these so-called degassing holes.”

If sensitive signals must be surrounded by ground, those signals may ride over and under such a plane, with ground lines on either side of the signal line or bus to reduce loop inductance. “Let’s say you have 10 traces that have four-micron pitch,” said Kelly. “You would want a big ground plane under them. And probably the best you can do in that scenario is to ‘waffle’ all the return currents. They don’t get to go directly under the traces because they have to walk through the waffle, but it’s still decent.”

Fig. 1: Ground plane with degassing holes. Ground planes can be placed under or over the signals they protect (or both). Blue arrows trace the rough current return paths for each red signal path below the plane. For two of the signals, the return currents must weave around the degassing holes. Source: Bryon Moyer/Semiconductor Engineering

An alternative to relying on the interposer for maintaining a clean signal is to assign that task to the package substrate. In this case, the goal is to get the signal off the interposer as quickly as possible using through-silicon vias (TSVs), which simply drops the signal directly down to the substrate.

“The industry is working on reducing the thickness of the TSV,” said Wang. “Then we directly transmit the signal to an ABF substrate. They have thicker metal lines and can create much better impedance than dense interposers.”

This only works for signals that would be driving an external ball on the package, however, even though technically one could run a TSV down to the substrate and another TSV back up to the interposer. The benefit of the interposer is to allow signals with dimensions below what’s possible on the substrate, so relying on the substrate for signals that would otherwise remain on the interposer would make little sense.

An active photonic interposer
Photonics is gradually moving farther into the data center, addressing orders of magnitude shorter distances than the kilometers that used to be the sole optical realm. On a data-center switch, single fibers can operate with multiple wavelengths and modes to provide greater bandwidth, while expanding the number of ports that can be served (known as the radix).

Traditional photonic connections to servers consist of pluggable optics that convert optical to electrical at the server edge, moving the signals over electrical SerDes for the remaining distance. But newer efforts focus on running optical signals all the way into packages to save power by eliminating the SerDes lines. That brings optical signals into the box, where they are converted to electrical for computation, and then back to optical for transmission over fiber. Lightmatter’s Passage, which is effectively a photonic interposer, can implement this within a package.

“The industry has been trying to bring optics closer to the chip for various reasons,” said Steve Klinger, vice president of product at Lightmatter. “The ultimate solution is where the optics are integrated into the package, and the ultimate extreme of that is where the optics are actually in the interposer and the chips are integrated directly on top of the photonic chip.”

Passage combines CMOS and silicon photonics in an interposer that can handle the conversion between optical and electrical domains, as well as circuits to control the photonics. It’s thus an active interposer. Electrical chips can be mounted as with any other interposer, but the interposer itself then converts the electrical signals to optical, routing them through waveguides and converting back to electrical when delivering a signal to a chip. That means the chips placed on the interposer require no electrical-to-photonic converters.

“[A die mounted on Passage] does have SerDes, but they can be very short-reach because they’re driving just a tiny electrical distance,” noted Klinger. “You don’t have to put the SerDes circuits all on the edge of the chip. They can be anywhere in the area of the chip. That’s one of the advantages of doing this integration. You eliminate that shoreline bottleneck.”

Light behaves much differently than electrons, and photonic “circuits” aren’t truly circuits in that they don’t obey conservation laws such as Kirchhoff’s Laws. Return currents aren’t a thing. So waveguides can run much farther with little loss. Lightmatter has built interposers as big as eight reticles, with waveguides experiencing little loss at the stitching points.

Little change in sight
The limitations of metal lines in interposers are fundamental and unlikely to change. For designs with lower signal density, it may be possible to improve the situation by widening lines, but that will be tough when very wide buses must cross from chip to chip.

Thickness is another possible parameter to explore. “On silicon interposers, we believe metal equal to or thicker than 2 micron can solve this kind of problem,” said Wang.

Ultimately, this becomes an exercise in placing components in a manner that keeps lines short. And all high-speed signals must have their signal integrity thoroughly analyzed to ensure that those signals can truly operate effectively at that high speed. Those checks should include not only the portion of the line running on the interposer, but every component along the path, including balls and bumps. Any ground planes must be modeled in that simulation.

Photonics can provide a longer-term way out of these line-length limitations, but it’s not in high-volume production yet, and it will be some time (if ever) before photonics can serve for lines that are long by electrical standards but extremely short by optical standards.

Related Reading
Why There Are Still No Commercial 3D-ICs
More than Moore is off to a good start, but the next steps are a lot more difficult.
3.5D: The Great Compromise
Pros and cons of a middle-ground chiplet assembly that combines 2.5D and 3D-IC.
Co-Packaged Optics Reaches Power Efficiency Tipping Point
But blazing fast data speeds come with significant manufacturing challenges.