One of things I have learned is that you cannot understand everything associated with designing and making of semiconductors. The physics associated with the manufacturing process, while highly interesting, involves a lot more material science than I care to learn. But for others this is the core of their business and I learned a lot more about it when I spoke to David Fried, the CTO of Coventor.
Now, while I cannot understand many of the details behind their product, the rational for it is blazingly clear. Let me start in the land of functional verification. We do simulation for a number of reasons. The first is that it is too expensive to try something out in silicon before you have a reasonable confidence that it will work. The second is that once you have created the device, you have limited visibility into what is going on inside the chip. So, simulation serves both purposes. First you can ensure that you only go to silicon when you are confident enough that it will work and secondarily, the simulation provides you much more visibility into what is happening so that debug is a lot simpler.
What Coventor does it to simulate the fabrication process and they have just done a major upgrade to their SEMulator3D product. Let me start by going over some of the economics associated with this tool and then I will attempt to explain what they have just introduced.
When a fab is about to start a new process, they have a lot of equipment to buy and a learning to do. It is estimated that a modern fab costs between $5B and $10B and to start a new process on this equipment will take an additional $2B. So there is a huge investment here. Today, there are many new technologies that a fab has to learn about including FinFETs, self-aligned interconnect, 3D Flash, TSVs, multi-patterning, 450mm wafers and possibly EUV.
David explained that a fab has to learn about each new process and the traditional way to do that is to create many experiments, run it through the fab and see what results you get. From these experiments they can learn about the electrical and physical properties of the process, its sensitivity to variations etc. Each of these experiment cycles takes about 3 months and costs $50M. David told me that it not untypical for a new process to take 15 of these cycles before the process has been characterized enough to be able to start accepting designs from customers. At the beginning of the process, smaller runs are likely to concentrate on fundamental issues while towards the back end of the process, more complex examples will be used. If one of these is lost then it produces a three month delay.
What SEMulator3D does, is to virtualize this process. A simulator can execute the process by building a high resolution model at each stage of production. The resolution is in the order of 5 Angstoms. Once the model is built, analysis can be performed on it, either visually to see what would have been built or analytically to see how it would behave. As a comparison, one of these simulations would take two and a half days using 512 CPUs and in that time would have conducted 150 5-way experiments. Now, these simulations cannot replace all of the trial runs, but they can ensure that each run performed is likely to produce exactly the results desired. It thus becomes possible to significantly speed up the process characterization and enable to fab to start taking designs a lot faster.
So, what is new? In the past SEMulator3D relied on voxels for modeling. These are just like pixels, only in three dimensions and could use many of the same types of algorithms as computer graphics. They utilized integer math and so these models could be built fast. Unfortunately as dimensions have become smaller and for some of the physical affects now going on inside the chip, it could not always produce results that were accurate enough. So Coventor has added a second form of modeling that they call Surface Evolution. This produces smooth curves rather than a pixelated model but the models rely on floating point math and so take longer to construct. However, the results are more accurate. SEMulator3D can use both of these to produce a high-speed behavioral model with advanced physics driven modeling.