Why a Tighter Flatness Spec Can Double Cold Plate Cost
Summary
A change from 0.025 mm to 0.015 mm looks minor on a cold plate drawing. On the manufacturing floor, it can mean a secondary finishing operation and a substantial increase in the cost of the contact face. But tighter flatness won’t necessarily improve cooling if the contact face’s surface texture isn’t properly controlled. Even a plate that meets its flatness tolerance can introduce unwanted thermal resistance. Compare machining capabilities and costs, identify the tolerances that trigger additional finishing operations, and learn what to specify on a cold plate drawing to meet thermal requirements without adding unnecessary manufacturing cost.
KEY TAKEAWAYS
ARTICLE
One of the weakest links in a direct-to-chip cooling loop is also one of the thinnest. It’s a layer of thermal interface material (TIM), typically 25 to 75 microns thick, between the processor and the cold plate.
TIM conducts a few W/m·K. The copper on either side of it conducts about 388. Every extra micron of TIM adds resistance to the heat path, and the cold plate’s flatness and surface finish decide how many microns end up there.
That makes the contact face one of the most important surfaces on the part, and one of the easiest to overspecify.
What Decides How Thick the TIM Bondline Gets
TIM has one job. It fills the microscopic gaps between the processor package and the cold plate so heat doesn’t have to cross air, which conducts even worse. It isn’t meant to make up for a contact face that’s out of flat.
On a well-specified interface, surface finish, TIM properties, and clamping conditions help determine the bondline. The TIM fills the peaks and valleys of the machined surface and little else, so the bondline stays in that 25 to 75 micron range.
On a poorly specified one, flatness error sets it. Wherever the face bows or dishes, TIM fills the extra distance, and the bondline can grow to several times what roughness alone requires. A plate dished 0.040 mm adds 40 microns of TIM in the region carrying peak heat flux.
That’s how a cold plate can meet its drawing and still let the processor throttle. Flatness checks out, no surface finish was specified, and under the specified clamping load, the processor drops to 80% of its rated clock speed.
Controlling the bondline means specifying flatness and surface finish together and knowing what each machining process can hold.
Cold Plate Flatness and Surface Finish by Machining Process
Each machining process has a natural range for flatness and surface finish. Specify inside that range and the face comes off in one operation. Specify past it and the part needs a different process.
| Process | As-Machined Flatness | Typical Ra |
|---|---|---|
| Standard CNC Milling | 0.050 to 0.080 mm | 1.6 to 3.2 µm |
| Precision CNC Milling | 0.025 to 0.050 mm | 0.8 to 1.6 µm |
| Single-Point Fly Cutting | 0.015 to 0.030 mm | 0.4 to 0.8 µm |
| Surface Grinding | 0.010 to 0.025 mm | 0.2 to 0.8 µm |
| Lapping | 0.003 to 0.010 mm | 0.05 to 0.4 µm |
Extruded and Brazed Cold Plates Start Farther from Flat
Extruded and brazed plates need facing before they can hold any of the ranges above. Their starting flatness is a condition to plan around, not a specification to call out.
| Starting Condition | Flatness Before Facing | Typical Ra After Facing |
|---|---|---|
| As-Extruded | 0.050 to 0.100 mm | 1.6 to 3.2 µm |
| As-Brazed | 0.100 to 0.250 mm | Set by the facing process |
Brazing makes the biggest difference. A vacuum braze cycle can relieve residual stress in the base, so the plate may leave the furnace in a different shape than it entered, sometimes with flatness variation an order of magnitude greater than the drawing allows.
The solution is sequencing. Machine the mounting face after brazing and put that requirement on the drawing as a process note.
How Tighter Flatness Changes the Process and the Price
Flatness cost doesn’t rise evenly with each tighter tolerance. It rises gradually while the same process can do the job, then jumps when the part needs a new operation.
Flatness cost doesn’t rise evenly with each tighter tolerance. It rises gradually while the same process can do the job, then jumps when the part needs a new operation.
That’s why two tolerances that look close on a drawing can price very differently. A callout of 0.020 mm can often stay within the primary CNC operation. A callout of 0.012 mm may require a separate finishing process, adding setup, fixturing, and inspection costs. If the thermal model meets requirements at 0.025 mm, that’s the number to specify. Going down to 0.015 mm usually adds more cost than cooling.
Before tightening a flatness tolerance, ask:
Surface Finish: The Other Half of the Interface
Flatness and surface finish are separate callouts, and a contact face needs both. Get either one wrong and the bondline suffers, just in different ways.
Surface roughness affects how well TIM fills microscopic peaks and valleys. If those irregularities aren’t adequately filled, voids increase thermal resistance. Flatness error creates a different problem: a bowed or dished contact face requires a thicker TIM bondline to bridge the uneven surface. Both conditions can raise junction temperature, even when the other specification passes inspection.
For many cold plate contact faces, 0.4 to 0.8 µm Ra is a useful working range, depending on the TIM and contact pressure. Finishing below 0.4 µm Ra may add cost with limited thermal benefit. Above about 1.6 µm Ra, surface texture can become a greater influence on bondline thickness, limiting the benefit of tighter flatness alone.
Why Ra Alone Isn’t Enough
Ra averages surface roughness into a single number. But two surfaces with the same Ra can have very different peak-to-valley profiles, which affect how TIM fills the interface.
Rz characterizes peak-to-valley roughness, providing additional information about the surface profile that Ra alone doesn’t capture. Few cold plate drawings call it out. For critical die interfaces, consider specifying both Ra and Rz. An Rz value approximately four to six times Ra can serve as a preliminary check, depending on the surface profile and measurement standard.
The direction of the machining marks matters too, and it’s rarely specified. Radial machining marks may provide paths for TIM to migrate outward during thermal cycling. Circumferential marks create a different surface pattern that may reduce that tendency. On a face that will see thousands of thermal cycles, that difference counts, and specifying lay direction costs nothing.
Why Large AI Packages Make Flatness Harder to Hold
Flatness tolerances haven’t tightened much as AI packages have grown. The distance they have to cover has.
Holding 0.025 mm across a 30 mm die works out to 1 part in 1,200. Across a 70 mm die, it’s 1 part in 2,800. Everything that pulls a face out of flat, including residual stress relaxation, heat from the cut, and fixture distortion, now acts over more than twice the distance.
What to Specify on a Cold Plate Contact Face
Specify the Interface, Not Just the Part
MES machines cold plate contact faces in copper and aluminum across 8 countries. We verify parts by CMM at the source, on the parts that ship, not just on a first article.
Every drawing gets an engineering review before a tool is cut. On a die interface, that review usually turns up a flatness zone that was never defined.
Send Us Your Cold Plate Drawing
Share your drawing, die footprint, and thermal requirements. An MES engineer will tell you what the contact face can hold and what it will cost.
Frequently Asked Questions About Cold Plate Flatness and Surface Finish
What flatness should a cold plate hold at the die interface?
For most direct-to-chip applications, 0.025 mm across the die contact footprint is a good working specification. A controlled CNC operation can usually hold it in a single setup. Below about 0.015 mm, a secondary operation such as grinding or dedicated fly cutting is often required, significantly increasing manufacturing cost. The flatness tolerance should reflect what the thermal model needs.
Why does cold plate flatness affect processor performance?
Flatness error thickens the thermal interface material bondline across part of the die. TIM conducts a few W/m·K, compared with about 388 W/m·K for copper, so every added micron raises thermal resistance. A plate dished 0.040 mm can add up to 40 microns of TIM locally, potentially in a region carrying peak heat flux.
What surface finish should a cold plate contact face have?
For many cold plate contact faces, 0.4 to 0.8 µm Ra is a useful working range, depending on the TIM and contact pressure. Finishing below 0.4 µm Ra may add cost with limited thermal benefit. Above about 1.6 µm Ra, surface texture can become a greater influence on bondline thickness, limiting the benefit of tighter flatness alone.
What’s the difference between Ra and Rz on a cold plate?
Ra averages surface roughness into one number, so surfaces with the same Ra can have different peak-to-valley profiles. Rz characterizes those variations, providing additional information about how the surface may interact with thermal interface material. For critical die interfaces, consider specifying both. An Rz value approximately four to six times Ra can serve as a preliminary check, depending on the surface profile and measurement standard.
Why does a brazed cold plate need machining after brazing?
A vacuum braze cycle heats the entire assembly and can relieve residual stress in the base, changing the plate’s shape during processing. As-brazed flatness commonly runs 0.100 to 0.250 mm, an order of magnitude worse than a typical specification. Machining the mounting face after brazing should be called out on the drawing as a process note.
Why is flatness harder to hold on large AI packages?
The tolerance usually stays the same while the span grows. Holding 0.025 mm across 30 mm is 1 part in 1,200, and across 70 mm it’s 1 part in 2,800. Residual stress relaxation, heat from the cut, and fixture distortion all act over more than twice the distance.
