Design for manufacturability

Traditionally, in the pre-nanometer era, DFM consisted of a set of different methodologies trying to enforce some soft (recommended) design rules regarding the shapes and polygons of the physical layout of an integrated circuit. These DFM methodologies worked primarily at the full chip level. Additionally, worst-case simulations at different levels of abstraction were applied to minimize the impact of process variations on performance and other types of parametric yield loss. All these different types of worst-case simulations were essentially based on a base set of worst-case (or corner) SPICE device parameter files that were intended to represent the variability of transistor performance over the full range of variation in a fabrication process. Additionally, SPICE models should have mismatches built into them for analog circuit simulations. Many mismatches are size and orientation dependent, which can be well modeled. Always "copy exactly" when doing analog layouts as many mismatches are not well understood or controlled (i.e. if one device has North to South current flow, all matched devices should have North to South current flow).

The most important yield loss models (YLMs) for VLSI ICs can be classified into several categories based on their nature.

• Functional yield loss is still the dominant factor and is caused by mechanisms such as misprocessing (e.g., equipment-related problems), systematic effects such as printability or planarization problems, and purely random defects.
• High-performance products may exhibit parametric design marginalities caused by either process fluctuations or environmental factors (such as supply voltage or temperature).
• The test-related yield losses, which are caused by incorrect testing, can also play a significant role.

After understanding the causes of yield loss, the next step is to make the design as resistant as possible. Techniques used for this include:

• Substituting higher yield cells where permitted by timing, power, and routability.
• Changing the spacing and width of the interconnect wires, where possible
• Optimizing the amount of redundancy in internal memories.
• Substituting fault tolerant (redundant) vias in a design where possible

All of these require a detailed understanding of yield loss mechanisms, since these changes trade off against one another. For example, introducing redundant vias will reduce the chance of via problems, but increase the chance of unwanted shorts. Whether this is good idea, therefore, depends on the details of the yield loss models and the characteristics of the particular design.

The objective is to design for lower cost. The cost is driven by time, so the design must minimize the time required to not just machine (remove the material), but also the set-up time of the CNC machine, NC programming, fixturing and many other activities that are dependent on the complexity and size of the part.

Unless a 4th and/or 5th axis is used, a CNC can only approach the part from a single direction. One side must be machined at a time (called an operation or op). Then the part must be flipped from side to side to machine all of the features. The geometry of the features dictates whether the part must be flipped over or not. The more ops (flip of the part), the more expensive the part because it incurs substantial set-up and load/unload time.

Each operation (flip of the part) has set-up time, machine time, time to load/unload tools, time to load/unload parts, and time to create the NC program for each operation. If a part has only 1 operation, then parts only have to be loaded/unloaded once. If it has 5 operations, then load/unload time is significant.

The low hanging fruit is minimizing the number of operations (flip of the part) to create significant savings. For example, it may take only 2 minutes to machine the face of a small part, but it will take an hour to set the machine up to do it. Or, if there are 5 operations at 1.5 hours each, but only 30 minutes total machine time, then 7.5 hours is charged for just 30 minutes of machining.[2]

Lastly, the volume (number of parts to machine) plays a critical role in amortizing the set-up time, programming time and other activities into the cost of the part. In the example above, the part in quantities of 10 could cost 7–10 times the cost in quantities of 100.

Typically, the law of diminishing returns presents itself at volumes of 100–300 because set-up times, custom tooling and fixturing can be amortized into the noise.[3]

The most easily machined types of metals include aluminum, brass, and softer metals. As materials get harder, denser and stronger, such as steel, stainless steel, titanium, and exotic alloys, they become much harder to machine and take much longer, thus being less manufacturable. Most types of plastic are easy to machine, although additions of fiberglass or carbon fiber can reduce the machinability. Plastics that are particularly soft and gummy may have machinability problems of their own.

Metals come in all forms. In the case of aluminum as an example, bar stock and plate are the two most common forms from which machined parts are made. The size and shape of the component may determine which form of material must be used. It is common for engineering drawings to specify one form over the other. Bar stock is generally close to 1/2 of the cost of plate on a per pound basis. So although the material form isn't directly related to the geometry of the component, cost can be removed at the design stage by specifying the least expensive form of the material.

A significant contributing factor to the cost of a machined component is the geometric tolerance to which the features must be made. The tighter the tolerance required, the more expensive the component will be to machine. When designing, specify the loosest tolerance that will serve the function of the component. Tolerances must be specified on a feature by feature basis. There are creative ways to engineer components with lower tolerances that still perform as well as ones with higher tolerances.

As machining is a subtractive process, the time to remove the material is a major factor in determining the machining cost. The volume and shape of the material to be removed as well as how fast the tools can be fed will determine the machining time. When using milling cutters, the strength and stiffness of the tool which is determined in part by the length to diameter ratio of the tool will play the largest role in determining that speed. The shorter the tool is relative to its diameter the faster it can be fed through the material. A ratio of 3:1 (L:D) or under is optimum.[4] If that ratio cannot be achieved, a solution like this depicted here can be used.[5] For holes, the length to diameter ratio of the tools are less critical, but should still be kept under 10:1.

There are many other types of features which are more or less expensive to machine. Generally chamfers cost less to machine than radii on outer horizontal edges. 3D interpolation is used to create radii on edges that are not on the same plane which incur 10X the cost.[6] Undercuts are more expensive to machine. Features that require smaller tools, regardless of L:D ratio, are more expensive.