M9 copper clad laminate: the 224G SerDes material set
M9 copper clad laminate is the lowest-loss copper clad laminate in volume production today. Dielectric loss runs 0.0005 to 0.0009 at 10 GHz. Four material families carry that number: a hydrocarbon-dominant resin, quartz fiber cloth instead of glass cloth, HVLP4 or HVLP5 ultra-low profile copper foil, and low-alpha spherical silica filler. Swap any one of them and the stack drops back toward M8.
Why M9 exists at all: the loss problem at 224 Gbps
Every serial link loses amplitude as it travels. Higher frequency means faster loss, and a shorter distance before the receiver cannot tell a one from a zero. Equalization buys margin back. It also costs power and design work, and eventually it runs out.
Move a design to 224 Gbps per lane and the channel stops being free. A backplane trace is not a short hop, and an AI server backplane material has to survive the whole run. On a 78 layer or 87 layer AI server board it runs the full depth of the chassis, so material loss per unit length gets multiplied by a very long path. At that point no layout trick helps. The laminate loses less, or the link fails.
M9 answers that constraint. It is a target with three parts, not a single specification:
- Signal loss low enough to survive a long backplane run. Tens of centimeters of board, and the waveform still has to be readable at the far end.
- Dielectric behavior stable across the full operating temperature range. An AI server sits at high load for weeks. An impedance that drifts with temperature will break the link.
- Thermal expansion matched to the silicon. Third-generation 3 nanometer class silicon and the laminate have to move together, or the board warps and joints crack.
Those three requirements pull against each other, and everything else in the M9 material system follows from them.
The four material families inside M9 copper clad laminate
M9 rebuilds four families at once. It is not a formulation tweak. The table below is the whole article in compressed form: what each family contributes, what M9 specifies, and why the change was forced.
| Family | What it contributes | What M9 specifies | Why M9 forces the change |
| Resin | Sets the floor on dielectric loss | Hydrocarbon resin as the main phase with PPO as the auxiliary, compounded at roughly 2 to 1 | A PPO-dominant route cannot reach the M9 loss window |
| Reinforcement | Carries mechanical load and holds dimensional stability | High purity quartz fiber cloth, known as Q cloth | E-glass cloth brings polarizable ions and expands more with heat |
| Conductor | Carries the signal | HVLP4 or HVLP5 ultra-low profile copper foil, with reverse treated foil used in some builds | Surface roughness dominates conductor loss once frequency climbs |
| Filler | Trims z-axis expansion and thermal path | Low-alpha spherical silica with high sphericity | Expansion and heat removal both sit outside what the resin alone can control |
Choosing a 224G SerDes PCB material means choosing all four together, and the same applies to any high speed copper clad laminate above M8.
Two of the four are supply chain problems rather than chemistry problems. Quartz fiber cloth and HVLP5 foil come from a very small number of plants worldwide. That scarcity sets both the cost and the schedule of an M9 program.
The loss ladder: what separates M7, M8 and M9
Copper clad laminate is graded by dielectric loss. Higher numbers mean lower loss and longer reach. Three rungs matter right now.
| Grade | Typical Df at 10 GHz | What it is designed to carry | Where it stands |
| M7 | 0.002 to 0.003 | 800G switches and first-generation AI servers | Mature, still the volume workhorse |
| M8 | 0.0012 to 0.0015 | 224 Gbps links at their early stage, and 1.6T optical modules | In ramp |
| M9 | 0.0005 to 0.0009 | 224 Gbps and 448 Gbps SerDes, 78 to 87 layer backplanes, Rubin class compute platforms | The volume ceiling today |
CCL is the abbreviation the industry uses for copper clad laminate, and this ladder is the scale it grades on. An ultra low loss copper clad laminate means M8 or above today, and the phrase increasingly points at M9.
Each step down the ladder is roughly a factor of two in loss. That is not a comfortable margin. It is the difference between a link that closes with power to spare and one that needs an expensive retimer. And the Df figure only applies at the frequency it was measured. Loss rises with frequency, so a 10 GHz value is a screening number, not a design number for a 28 GHz channel.
Resin: the shift from a PPO backbone to a hydrocarbon backbone
The resin decides the dielectric floor, and M9 breaks with the two generations before it. The choice of hydrocarbon resin for PCB laminate sets the whole grade. M6 through M8 were built on PPO, or PPE, as the main phase with hydrocarbon resin as an additive. M9 inverts that. Hydrocarbon resin becomes the main phase, PPO becomes the auxiliary, and specialty functional additives close the gaps.
The main phase. A specialty hydrocarbon resin contains only carbon and hydrogen. No polar groups sit in the chain, so the molecules respond very little to an alternating field. That is what low dielectric loss means at the molecular level. Purity is pushed to 99.99 percent with impurities held below 0.01 percent. In practice the resin is compounded with PPO at roughly 2 to 1, because a pure hydrocarbon backbone is hard to process and bonds poorly.
The functional components. Low polarity is also the problem. A non-polar chain does not stick well to copper or to glass, and it does not crosslink easily. Bismaleimide, active ester and benzoxazine go in as crosslinkers to restore adhesion and cure behavior. Higher-end formulations add benzocyclobutene or a polyimide-modified cyanate ester to lift thermal stability further.
| Route | Molecular idea | Reported performance | What it buys |
| Flexible ether spacer with a symmetric triazine ring | Separates polar groups and keeps the backbone symmetric | Decomposition temperature 342 degrees Celsius with Df 0.0014 at 28 GHz | Thermal stability and low loss in the same backbone |
| Hydrocarbon resin with nano silica doping | Dilutes the polarizable fraction | Dk variation within plus or minus 0.01 from 2 to 110 GHz | A dielectric response that stays flat over a very wide band |
The resulting window is what makes M9 possible at all. Glass transition temperature sits at 250 degrees Celsius or higher, with premium formulations past 350 degrees Celsius. Decomposition temperature goes above 420 degrees Celsius. Dielectric loss can reach 0.0006 at 10 GHz. No epoxy system comes close.
Quartz fiber cloth: the gate that defines the M9 grade
Below M9, boards are reinforced with E-glass cloth or a low Dk glass cloth. M9 requires high purity quartz fiber cloth. This one substitution is the clearest line between generations.
| Property | What quartz cloth delivers | Why it matters |
| Dielectric performance | Dk around 2.3 and Df around 0.0009, reported values | Almost no additional loss is introduced by the reinforcement itself |
| Thermal expansion | Expansion coefficient very close to silicon | Sharp reduction in thermal stress and warpage risk on boards above 70 layers |
| Purity | Silicon dioxide content approaching 100 percent, no alkali metal ions | Removes the ion polarization loss that appears at high frequency |
A quartz fiber cloth CCL is not a drop-in upgrade, because the whole fabrication line moves with it.
The catch is manufacturing. Quartz fiber is extremely brittle. It breaks during drawing, twisting and weaving at rates that would shut down an E-glass line, and the process window is narrow. Equipment precision has to be an order of magnitude better than what E-glass needs. That is why quartz cloth is the most constrained link in the M9 chain, and why capacity rather than chemistry sets the pace of adoption.
Copper foil: HVLP4 and HVLP5 against the skin effect
High frequency current does not travel evenly through a conductor. It crowds into a thin layer near the surface, and how thin that layer is depends on frequency. Since the current only sees the outer skin, the shape of the copper surface sets the conductor loss.
M9 standardizes on HVLP4 and HVLP5 ultra-low profile copper foil. Reverse treated foil gets chosen for some builds. Within the HVLP copper foil family, HVLP4 copper foil and HVLP5 copper foil differ mainly in how tight the surface profile is.
| Foil class | Surface roughness | Where it is used | Effect on loss |
| Conventional foil | Ra around 1.5 micrometers | Commodity FR-4 boards | Rough peaks and valleys lengthen the current path |
| HVLP4 | Ra at or below 0.4 micrometers | M9 builds from 224 Gbps upward | Keeps the skin-effect path short and predictable |
| HVLP5 | Tighter than the HVLP4 window | M9 builds with the least loss budget | Same mechanism, less headroom consumed |
| Reverse treated foil | Between conventional and HVLP | Selected M9 builds | Trades some roughness for better peel strength |
Processing matters as much as the foil. Microvia filling electroplating and a controlled micro-etch, with etch depth held between 1.8 and 2.1 micrometers, balance a smooth profile against the peel strength the board still needs. The copper itself is high purity electrolytic copper, because impurities scatter the signal.
Spherical silica filler: a small fraction with a large effect
Filler has the smallest share of the formulation and the largest influence on reliability.
| Filler requirement | Target | Why it is specified |
| Sphericity | Above 95 percent | Controls packing and flow inside a thin prepreg |
| Particle size distribution | Narrow | Keeps the dielectric response predictable |
| Radioactivity | Low alpha emission | Prevents soft errors in adjacent silicon |
| Magnetic foreign matter | Very low | Protects signal integrity and device reliability |
| Base material | High purity spherical silicon dioxide | Low loss together with a lower z-axis expansion |
Filler does three jobs. It lowers the z-axis coefficient of thermal expansion, which improves dimensional stability. It trims Dk and Df further. It raises thermal conductivity, which relieves the heat load from high density packaging around an AI accelerator. It is also the reason drill bits die young, which is the next problem in the chain.
What M9 does to PCB fabrication
Quartz cloth, hard filler and non-polar hydrocarbon resin together overturn the processing behavior a shop learned on FR-4. M8 to M9 is a generational change in fabrication difficulty, not an incremental one. Drilling is where it hurts most.
Drilling: the step where M9 changes the economics
Quartz fiber has a Mohs hardness near 7. E-glass sits around 4.5. Add the abrasive effect of hard spherical silica and tool wear accelerates sharply.
| Substrate | Reinforcement hardness | Drill life per bit | What the shop sees |
| M7 | E-glass, Mohs around 4.5 | 500 to 1,000 holes | An established process |
| M8 | Low Dk glass cloth | 400 to 500 holes | Tool wear starts to matter |
| M9 | Quartz fiber, Mohs around 7 | 100 to 200 holes | 70 to 90 percent of tool life disappears |
The response has been a move to PCD diamond micro drills, which changes the cost structure of the whole operation.
| Item | Tungsten carbide bit | PCD diamond micro drill |
| Life on M9 | 100 to 200 holes | 8,000 to 15,000 holes |
| Unit price | Baseline | 5 to 10 times baseline |
| Hole wall quality | Rz at 35 micrometers or worse, with burrs and a risk of delamination | Ra held within 0.8 micrometers |
| What it demands | Existing parameters | Spindle speed and feed re-tuned from scratch |
Thickness compounds it. Past 78 layers a backplane reaches 8 to 10 millimeters thick, and the drill has to pass all the way through. Effective drill length moves from the traditional 5 millimeters to more than 10 millimeters. A long slender drill flexes more and breaks more often.
Add it up and a high layer count PCB in M9 costs 4 to 8 times the same drilling step on M7. It becomes the most expensive operation in the process.
Lamination: 87 layers and a 0.5 parts per thousand warpage target
Layer counts jump from 20 to 30 in the previous generation to 44 to 87. The pressed stack is thick, and getting heat and pressure through it evenly is a different problem from what a 30 layer board presents. Warpage control in lamination is what turns that into a schedule problem.
| Parameter | M7 class | M9 class | Consequence |
| Layer count | 20 to 30 | 44 to 87 | Stack thickness and registration complexity rise together |
| Press cycle | Baseline | About 30 percent longer | Capacity per press drops |
| Process | Conventional vacuum press | Stepped vacuum press with controlled ramp, pressure curve and cure | Otherwise the stack delaminates and dielectric properties vary across the panel |
| Warpage | Standard control | Held within 0.5 parts per thousand | Controls snap-through risk at reflow |
Warpage comes from the mismatch between quartz cloth and resin. Their expansion coefficients differ, so internal stress concentrates after pressing, and the difficulty scales faster than the layer count. Leading makers use symmetric stack design and multi-stage constant temperature dwell to hold warpage inside the target. Press capacity falls as a direct result.
Etching, metallization and layer registration
Three more steps change character with M9.
| Process step | Why M9 makes it harder | What the shop changes |
| Line etching | Smooth HVLP foil narrows the window for side etch control | Precision exposure and differential etching to hold features below 50 micrometers |
| Hole metallization | Hydrocarbon resin is non-polar, so electroless copper adhesion is weak | Dedicated hole wall roughening and activation steps, or the copper peels |
| Layer registration | More layers multiply the cumulative difference in shrinkage | Tolerance tightens from plus or minus 25 to plus or minus 15 micrometers, with shrinkage compensation models |
Layer registration tolerance is the one that scales with layer count.
Each of the three is solvable on its own. The difficulty is solving them at the same time, on the same panel, without disturbing the dielectric properties that justified M9 in the first place.
The M9 copper clad laminate supply chain: who makes what
There is no single M9 recipe. Each maker assembles its own combination, and the differences concentrate in the resin system and the reinforcement choice.
| Supplier | Material combination | Reported performance | Positioning |
| Panasonic, Megtron 9 and 10 | Q cloth with specialty hydrocarbon resin and ultra-low profile foil | Df 0.0008 at 10 GHz, Dk 3.0 | Validated across the Nvidia and HiSilicon platforms |
| Rogers, RO4835+ | Hydrocarbon resin system with a gradient cure hot pressing process | Z-axis CTE down to 12 ppm per degree Celsius | Built around warpage control |
| Sumitomo Electric, Isola | BCB or BT modified resin systems | Strong phase stability at high frequency | Aimed at phase-sensitive designs |
Upstream has moved fast. All four material segments now have domestic supply, and the notable change is that quartz cloth and hydrocarbon resin, the two hardest segments, are no longer purely imported.
| Material segment | Supplier | Status | Notes |
| Hydrocarbon resin | Dongcai Technology | Volume supply, the only domestic source at M9 grade | Covers the BMI, active ester, PPO and hydrocarbon series, and sits inside the Nvidia, Shengyi and Taiwan Union supply chains |
| Hydrocarbon resin | Shengquan Group | Pilot validation | Working with a research institute on AI-assisted reverse design of a new resin monomer, with Df reported 19 percent below the conventional route |
| Quartz fiber cloth | Feilihua | Volume supply, domestic exclusive and one of three producers worldwide | Silicon dioxide purity above 99.99 percent, certified by Panasonic and Nvidia, core supplier to Shengyi |
| Copper foil | Longyang Electronics | HVLP5 in volume production | Inside the domestic M9 chain |
| Copper foil | Tongguan Copper Foil | HVLP4 in volume supply | Inside the domestic M9 chain |
| Spherical silica | Lianrui New Materials | Volume supply of low-alpha grades | Supplying Rubin class backplanes and GB300 packaging |
On finished laminate the field is narrower.
| Maker | Status | Detail |
| Shengyi Technology | Stable volume production | First mainland laminate maker to pass Nvidia M9 certification and the third company worldwide to do so, with board level Dk spread held within plus or minus 0.012 |
| Nanya New Materials | Small batch ramp | Following Shengyi |
| Others | Development or pilot | Not yet at volume |
The shape of this chain matters more than any single name in it. Domestic supply has crossed from zero to one. What is happening now is the shift from one to many, and the constraint has moved from whether a material can be made to whether it can be made at the volume and consistency certification demands.
M9 is not a formula, it is a set of trade-offs
Read a specification table and the instinct is to want the lowest loss number on it. That instinct is what gets M9 programs into trouble. Every electrical gain in this system is paid for somewhere else.
- Non-polar hydrocarbon resin delivers the low loss, and it bonds badly. The crosslinkers that fix adhesion change the cure window and add cost.
- Quartz cloth removes reinforcement loss and matches silicon expansion, and it is brittle. Yield through weaving and capacity in the supply chain are the price.
- Smooth copper lowers conductor loss, and it lowers peel strength. That is why micro-etch depth is held inside a 1.8 to 2.1 micrometer window, and why some builds accept a rougher reverse treated foil instead.
- Low-alpha silica cuts z-axis expansion and improves heat removal. Its hardness is what destroys drill bits and drives the move to diamond tooling.
- A longer press cycle improves lamination quality. It also cuts press capacity by roughly 30 percent.
So the grade is not decided by whoever prints the lowest Df on a datasheet. It is decided by whoever manages the exchange between electrical performance, mechanical margin and process cost, panel after panel. Two laminates with identical loss figures can behave completely differently in a fabrication line. The difference shows up in yield, not in the datasheet.
There is also a hard limit on how far this goes. Loss can be lowered, but the material loss floor is one term in the channel budget. Connectors, vias, the package substrate and the die all consume margin too. Past a certain point another material improvement buys less than the same money spent on equalization or on a shorter channel.
What to specify when you buy M9 material
The practical checklist is short, and a single loss figure answers none of it.
| What you are specifying | Ask for | Why the number matters |
| Loss | Df at the frequency you will actually run, not a 10 GHz value | Loss rises with frequency, so a 10 GHz screening figure can be optimistic by a wide margin |
| Consistency | Dk spread across the panel and across temperature | A low average with wide spread still breaks the link budget |
| Thermal | Glass transition temperature and decomposition temperature | These set the reflow and thermal cycling margin |
| Expansion | Z-axis CTE | Warpage and via reliability follow from it |
| Conductor | Foil class and surface roughness | Conductor loss lives here, and so does peel strength |
| Reinforcement | Quartz cloth grade and weave | The hardest item in the chain to substitute or reschedule |
| Filler | Sphericity, alpha emission, magnetic content | Reliability of the finished assembly |
Limitations
This article maps publicly available industry information. It is not a datasheet, and several caveats apply.
- Loss, dielectric constant and expansion figures are typical or reported ranges. They are not guaranteed values for any specific product. Never copy them into a purchase specification or a design rule without confirming with the supplier.
- Dielectric properties depend on test frequency and test method. A value quoted at 10 GHz does not transfer to another frequency, and values published by different sources at different frequencies are not directly comparable. Reinforcement cloth data in particular is published inconsistently, so verify against a current supplier datasheet.
- Drill life, press cycle length and cost multipliers are shop floor observations. They depend on the stack, hole size, machine and tooling. Treat them as orders of magnitude.
- Certification status and supply positions change fast in this segment. What was true of a supplier when this was written may not hold later.
- Nothing here about any company is investment advice or a statement about current commercial performance.
Bottom line
M9 is not a better resin. It is four material families moved together, and the reason is a channel loss budget that a long backplane cannot absorb. The resin sets the dielectric floor. Quartz cloth removes reinforcement loss and matches silicon expansion. HVLP foil keeps the skin effect path clean. Low-alpha silica holds the stack together thermally. Every one of those choices takes something away, which is why drilling costs four to eight times the M7 figure and press cycles run about 30 percent longer. The grade belongs where a channel genuinely needs 224 Gbps or 448 Gbps reach.
FAQ
What is M9 copper clad laminate?
M9 copper clad laminate is the lowest-loss grade of copper clad laminate in volume production. It is built for 224 Gbps and 448 Gbps serial links, 78 to 87 layer AI server backplanes and Rubin class compute platforms. Four material families define it at once: a hydrocarbon-dominant resin, quartz fiber cloth, HVLP4 or HVLP5 copper foil and low-alpha spherical silica filler.
What dielectric loss does M9 copper clad laminate reach?
Typical dielectric loss for M9 sits between 0.0005 and 0.0009 at 10 GHz, against 0.0012 to 0.0015 for M8 and 0.002 to 0.003 for M7. Premium constructions report figures such as 0.0006 or 0.0008 at 10 GHz. Treat any single figure as frequency specific, because loss climbs as frequency rises. Always confirm the test frequency before comparing two laminates.
Why does M9 need quartz fiber cloth instead of glass cloth?
Glass cloth carries polarizable alkali ions and expands more with heat, so it adds both loss and mechanical stress. Quartz fiber cloth reports a dielectric constant near 2.3 and a loss tangent near 0.0009, with silicon dioxide content approaching 100 percent. Its very low expansion coefficient also matches silicon closely, which matters on boards above 70 layers.
What copper foil does M9 use?
M9 uses HVLP4 or HVLP5 ultra-low profile copper foil, with reverse treated foil selected for some builds. Surface roughness is held at or below 0.4 micrometers, against roughly 1.5 micrometers for conventional foil. Because high frequency current travels in a thin layer near the surface, a smoother profile directly cuts conductor loss.
Why is drilling the hardest step in M9 PCB fabrication?
Quartz fiber has a Mohs hardness near 7, against roughly 4.5 for E-glass, and the spherical silica filler is abrasive. Tungsten carbide bits that drill 500 to 1,000 holes in M7 manage only 100 to 200 holes in M9. PCD diamond micro drills restore life to 8,000 to 15,000 holes, at five to ten times the price.
How many layers do M9 boards have?
M9 backplanes run from 44 to 87 layers, up from 20 to 30 layers in the previous generation, and designs at 78 layers and above reach 8 to 10 millimeters thick. That thickness forces drill bits with an effective length beyond 10 millimeters and pushes warpage control to 0.5 parts per thousand. Long slender drills flex, so breakage risk rises with thickness.
Which companies make M9 copper clad laminate?
Only a small group ships M9 grade laminate. Panasonic and Rogers lead on the material side, with Sumitomo Electric and Isola close behind on phase-stable systems. Shengyi Technology is the only mainland maker in stable volume production and the third company worldwide to pass Nvidia M9 certification. Nanya New Materials is ramping.
Who supplies M9 hydrocarbon resin?
Dongcai Technology is the only domestic supplier shipping M9 grade electronic hydrocarbon resin in volume, and it covers the BMI, active ester, PPO and hydrocarbon series. It already sits inside the Nvidia, Shengyi and Taiwan Union supply chains. Shengquan Group has an M9 resin in pilot validation, built with AI-assisted monomer design.
Is quartz fiber cloth the bottleneck in the M9 supply chain?
Yes, for now. Quartz fiber is brittle, so breakage during drawing, twisting and weaving runs high, and the equipment precision required sits an order of magnitude above E-glass. One producer holds the domestic position and three companies worldwide can make M9 grade cloth at volume, which makes it the tightest link.
Does a lower Df always mean a better board?
No. Every electrical gain is paid for elsewhere. Non-polar hydrocarbon resin weakens adhesion, so crosslinkers are added and hole metallization needs extra steps. Smooth foil lowers loss and also lowers peel strength. Hard filler cuts expansion and destroys drill bits. The grade is decided by whoever manages that exchange, not by the datasheet.
What should you ask for when specifying M9 material?
Ask for dielectric loss at the frequency you will actually run, dielectric constant spread across the panel and across temperature, glass transition and decomposition temperatures, z-axis CTE, foil class and surface roughness, plus cloth grade and filler alpha emission. One low loss figure says very little about whether a build will pass.
Where does M9 stop being the right answer?
M9 carries a cost and yield penalty. Drilling alone costs four to eight times the M7 figure, press cycles run about 30 percent longer and layer registration tightens to plus or minus 15 micrometers. If the channel does not genuinely need 224 Gbps or 448 Gbps reach, an M8 stack usually gives a better return.
