One of the challenges with any new tool is to drive adoption. There really is a "he who goes first loses" dynamic to a new product. Any new product has potential maturity issues early on, so everyone is better off if they wait for other people to go first. Of course, the trouble with that attitude is that if everyone takes it, then the product never matures. Somebody has to go first. The dynamic is worse with a product addressing very advanced process nodes since when the code is being created and tested there are no test cases in the advanced nodes. Tempus, the new timing analysis tool (officially called the Tempus Timing Signoff Solution) has had an incredible adoption ramp. Today Cadence announced that they have had over 200 tapeouts for Tempus within the first two years since it was announced as the first of the "us" products. These are not small designs either, some are 100 million instances or more. That is a couple of tapeouts per week. There are actually two different aspects to "he who goes first loses". The first is that customers don't want to be one of the early tapeouts, they would rather wait. But salespeople are responsive to their customers and if the customers don't want a product due to immaturity then they won't sell it. After all, if there turns out to be problems, it reflects badly on them and ties up their application resources. One thing that has changed is that doing design at the most advanced nodes is much more of a partnership between the design group, Cadence, and the foundries. Development of solutions is much more concurrent than it used to be in the days when one process was pretty much like the last one, and when EDA tools for one node were not that different from the previous node except, perhaps, for capacity. Now each node seems to bring new aspects of the process that require major changes to the design tools (such as double patterning and FinFETs to name the two most obvious). These new aspects require much closer cooperation than when a process just required design rules and SPICE parameters, with the result that even if customers would rather wait in some theoretical sense, they cannot. They have to get their designs taped out. I talked to Ruben Molina to find out more. His history was that he ran marketing at ExtremeDA (which did timing analysis), which got sued and then acquired by Synopsys. So he went to Magma and became product manager for Tekton timing analysis. But Magma also got sued and acquired by Synopsys. So in 2012, he joined Cadence. I think he is safe this time. The challenge with Tempus was to rearchitect the timing analysis to do two things. Firstly, make it handle all the physics issues of modern processes. Secondly, to handle designs of the size they are, to upgrade and rewrite everything to be massively parallel, so that it can take advantage of dozens or hundreds of machines. The Tempus timing is also a core engine that underlies the entire "us" product family with tools such as Genus synthesis and Innovus physical design. The parallelization is at two levels. The most obvious is that different corners can be run on different machines, which is not especially hard to implement since different corners do not interact with each other. Harder is to partition large designs across large numbers of servers, which Tempus can also do. So it is a sort of two-dimensional parallelism. Tempus also has path-based analysis, which leads to better PPA due to reduced pessimism, which leads to minimizing unnecessary fixes, in particular fewer buffers being inserted. So what is responsible for this very fast adoption rate? First, there has been a lot of work done with the foundries to certify it for signoff. In particular, there are new effects at 16nm, 14nm, 10nm, and beyond concerned with variation. Cadence has been working with the Liberty TAB on Liberty Variation Format (LVF). In fact, even that is not sufficient since there are implications from non-Gaussian distributions (asymmetric bell curves), and next year there will be work with the foundries on how to model skewness.![]()
![]()
↧