Is NVIDIA Eyeing Teflon?

The arms race of AI servers has extended to PCB materials.

It is reported that NVIDIA is evaluating the use of PTFE (Polytetrafluoroethylene) in the high-speed PCBs of next-generation AI servers, with the potential application on the NVSwitch switch board of Rubin Ultra. The name PTFE may sound unfamiliar, but it is widely known by another name: Teflon, which is the coating used on non-stick cookware. What NVIDIA values is certainly not its non-stick property, but its low loss when transmitting high-frequency and high-speed signals.

There has been a recent change in the material roadmap of Rubin Ultra. SemiAnalysis released PTFE Pushing Copper Further on August 20, noting that NVIDIA still intends to use PTFE on the NVSwitch board, but the solution is no longer the "full replacement of fiberglass cloth" that was widely speculated in the market previously, but a hybrid architecture of PTFE Core + M9 Prepreg.

The shift from full replacement to a hybrid design indicates that the challenges lie not only in material properties. To reduce signal loss, high-layer PCBs must also be manufacturable and scalable for mass production, which is the real hurdle on this path.

Why Use Teflon?

As AI clusters continue to grow larger, the volume of data that needs to be exchanged between GPUs keeps increasing, and the speed of high-speed interconnections such as NVLink is also rising. No matter how fast the chips calculate, if data cannot be transmitted stably between GPUs, switching chips and other hardware, the efficiency of the entire system will not improve. At the rate of 200Gbps per lane, the signal loss on PCB traces can no longer be ignored.

What exactly is the frequency level? A 200Gbps lane generally uses PAM4 modulation, with a Nyquist frequency of around 50GHz. The higher the signal frequency, the greater the loss on the PCB, making the selection of board materials even more critical.

Loss on PCBs mainly comes from two sources. The first is conductor loss: at high frequencies, current crowds to the surface of copper wires, which is the skin effect. At 50GHz, the skin depth is only a few tenths of a micrometer, so the surface roughness of copper foil directly affects the loss, which is why low-roughness copper foil (HVLP) is gaining increasing attention. The second is dielectric loss: the insulating medium surrounding the copper wire absorbs part of the electromagnetic energy, and this effect becomes more obvious at higher frequencies. As PCB traces in AI servers become more complex, the number of layers increases, and the path signals travel gets longer, the electrical performance of the board material itself becomes an unavoidable variable in the design process.

In recent years, high-speed copper clad laminates have been evolving towards low dielectric constant and low dielectric loss: from traditional epoxy resin, to PPO/PPE, Hydrocarbon, and then to high-end materials such as M8 and M9, all aiming to make high-speed signal transmission smoother. PTFE has drawn attention because its advantages in low loss are more prominent.

Data from SemiAnalysis shows that PTFE has a dielectric constant Dk of about 2.1 at 10GHz, and a dielectric loss factor Df of about 0.0003, which is lower than that of common PPO/PPE and Hydrocarbon resins. Dk reflects the degree of the material's response to the electric field, and Df reflects how much electromagnetic energy is lost in the dielectric. For high-speed PCBs, the lower the Df, the more energy remains after the signal is transmitted. It should be noted that this is the data at 10GHz, while the actual operating frequency band is much higher. One characteristic of PTFE is that its Df changes very little with frequency, remaining at a very low level even at tens of GHz, so its advantages in the high-frequency band will only be more obvious than at 10GHz.

Another attraction of PTFE is related to fiberglass cloth. Conventional PCB substrates are made of resin and fiberglass cloth. The dielectric constants of fiberglass and resin are different, and the fiberglass weave is not uniform. If the signal trace happens to be on the fiberglass bundle or the resin area, the local Dk will be different, which is known as the fiberglass effect. If the two lines of a differential pair pass through different areas, their propagation speeds will differ, resulting in timing skew, which becomes more problematic at higher rates. The PTFE core layer has very little or even no fiberglass content, so this problem is naturally much less severe. Therefore, the previous market speculation of "full replacement of fiberglass cloth" was motivated by more than just low Df.

This is also the reason why NVIDIA wants to use PTFE on the NVSwitch board. Rubin Ultra needs to further expand the scale of GPU interconnection, the switching chips have to process a large amount of data, and the transmission rate of the switch board and mid-board also needs to be higher. Wccftech mentions that some high-end PCBs already have more than 70 layers, and even with top-tier M9, they may not be able to support future signal requirements. Since the existing material system may not be sufficient, people have to turn to materials with lower dielectric loss.

The Challenges Lie in Manufacturing

The electrical performance of PTFE is not the biggest problem; the trouble lies in manufacturing.

The previous market speculation that Rubin Ultra may use PTFE to fully replace fiberglass cloth ignores a very practical problem: no matter how good the electrical performance of the material is, it has to go through processes such as drilling, lamination, and electroplating to finally form a reliable multi-layer board.

PTFE has very stable chemical properties and low surface energy, making it difficult to bond with other materials. This makes it an excellent low-loss dielectric, but also gives PCB manufacturers a headache.

First, let's talk about drilling. PTFE is relatively soft and thermoplastic. When friction generates heat during drilling, it tends to soften, resulting in residual wall material and burrs on the hole wall. If the material is mixed with fiberglass or other fillers, the hardness is uneven, making drilling more difficult, and the hole wall quality and dimensional accuracy will be affected. High-layer boards also need backdrilling, which drills off the excess copper stub in the through hole to avoid signal reflection caused by it at high frequencies. Backdrilling has very high requirements for depth and position accuracy. Once the drilling quality is unstable, the backdrilling process will also be affected.

After drilling, metallization is more troublesome. Traditional PCBs need desmearing, chemical copper deposition, and electroplating after drilling to reliably connect all layers. However, the surface of PTFE is too stable, the permanganate desmearing process commonly used for ordinary PCBs is basically ineffective for it, and copper is also very difficult to attach directly. Manufacturers need to use special chemical treatment or plasma treatment to activate the surface and increase the surface energy, so that the copper layer can adhere better.

Then there is multi-layer lamination. The melting point of PTFE is around 327°C, which is much higher than the lamination temperature commonly used for epoxy systems. The processing conditions are very different. When different materials are mixed in the same board, the lamination process has to be re-adjusted. The bonding between PTFE and other materials is also a difficulty, which usually needs to be solved by using special bonding sheets. In addition, PTFE is relatively soft and will deform slowly under pressure (cold flow), making the pressure and thickness control during lamination more difficult.

There is also the problem of thermal expansion. The Z-axis thermal expansion coefficient of pure PTFE is very high, far higher than the approximately 17ppm/°C of copper. It will decrease after adding fillers, but there is still a gap. During high-temperature lamination and repeated thermal cycles, the dimensional change between layers will affect the reliability of through holes, and in severe cases, it will cause fatigue cracking of the copper layer. The thicker the board and the more layers it has, the more prominent this problem becomes.

These problems are difficult to solve on ordinary high-frequency PCBs, and they will only become more challenging when applied to AI server switch boards with dozens of layers that need to operate stably for a long time.

Therefore, the hybrid solution disclosed by SemiAnalysis seems more reasonable. Different from the previously rumored all-fiberglass-free solution, the PTFE Core + M9 Prepreg design uses PTFE only in some core layers, while the other layers still use M9-grade prepreg. PTFE is responsible for reducing the dielectric loss on the key high-speed signal paths, M9 provides mechanical support, and the processing technology is relatively mature, so the manufacturing difficulty of the entire board is reduced.

In a high-layer PCB, not every layer needs to undertake the same high-speed transmission task. Placing PTFE on the key layers with the highest loss requirements and using more mature M9 for the rest not only leverages the electrical advantages of PTFE, but also avoids the manufacturing difficulties of using PTFE for the entire board. Of course, the hybrid stack-up also has its cost: the two materials have different thermal expansion and contraction properties. If the stack-up design is asymmetric, the board is prone to warpage, which is an aspect that needs special control during design.

Therefore, this adjustment to the material roadmap of Rubin Ultra is less of a concession for the PTFE solution, but more of a more pragmatic mindset: instead of pursuing the use of only one material for the entire board, materials are allocated according to the functions and signal requirements of each layer.

Pushing Copper Interconnection One Step Further

The significance of PTFE goes far beyond replacing a PCB material. SemiAnalysis named its report PTFE Pushing Copper Further, which has pointed out the larger underlying industrial logic: while optical interconnection is advancing, the industry still wants to extend the use range of copper interconnection as much as possible.

The interconnection inside AI clusters is changing. As the number of GPUs and the scale of clusters increase, traditional copper connections are facing growing pressure in terms of transmission distance, signal loss and bandwidth, so optical interconnection has attracted attention, and CPO is also regarded as an important direction for future high-speed interconnection.

However, copper and optics are not a simple substitution relationship.

The Vera Rubin Ultra NVL576 system announced by NVIDIA integrates 8 racks into a 576-GPU NVLink Domain, with each rack equipped with 72 Rubin Ultra GPUs, using both copper connections and direct optical connections. It can be seen that the future AI Scale-up network will not rely on only one transmission technology, but let copper and optics play their respective roles according to the requirements of distance, bandwidth and power consumption.

For short-distance connections, copper still has advantages in terms of cost, power consumption and manufacturing maturity. As long as the signal loss can continue to be reduced, there is still room for improvement in the speed of copper. This is where the value of PTFE lies: it is not intended to replace optical interconnection, but to improve the dielectric performance of PCBs, so that high-speed electrical signals can travel farther and more stably in copper lines.

PTFE, M9/M10, Low-Dk fiberglass, Q-cloth, low-roughness copper foil, these material systems are different, but they point to the same goal: to reduce the loss during high-speed signal transmission. The dielectric loss and conductor loss mentioned earlier correspond exactly to these two types of materials: PTFE, M9/M10, Low-Dk fiberglass and Q-cloth reduce dielectric loss, while low-roughness copper foil reduces conductor loss.

This also means that the competition for materials for high-end PCBs in the future will probably not be dominated by a single material, but will be matched according to the type of board and the signal requirements of each layer. High-speed switch boards and mid-boards may have an additional option of PTFE, while mature low-loss materials still have their place in other computing boards and general high-speed PCBs.

Of course, whether PTFE can really be adopted in the mass production design of Rubin Ultra still awaits final verification. Good electrical performance alone is not enough; drilling, lamination, metallization, thermal cycle reliability and overall yield must all meet the standards.

From GPUs, HBM to switching chips, and now to PCB substrates, the upgrading of AI infrastructure is digging deeper and deeper into the underlying layers. NVIDIA's focus on Teflon is ostensibly to find a new low-loss material, but in essence to gain more room for next-generation high-speed interconnection. Data transmission between chips is becoming increasingly critical, and even the circuit boards that carry data are involved in technological competition.

This article is from the WeChat Official Account "Semiconductor Industry Observation" (ID: icbank), written by the editorial department, and published with authorization from 36Kr.