Enough of the sideshows - it's time for some real advancement in functional verification!

Brian Bailey
EETimes (5/8/2012 9:23 AM EDT)

Abstract

Functional verification has become the congestion point for many designs and the bane of many companies trying to sell their IP. While it may appear as though there have been a plethora of new products over the past decade, I have often said that many of those so called advancements were superficial and that they were delaying the inevitable. Now, it is time for me to stop saying what is wrong and to start talking about a better direction for the future that solves many of the fundamentals problems. In this paper, I will review some of my past comments, define what is missing from the existing verification environments and, finally, discuss a way in which many of those problems can be solved or, at the very least, improved.

Introduction

Constrained random test pattern generation was hailed as a great advance over the previous mainstay of functional verification - directed testing. You will not find me disagreeing with that statement in general, but not for the reasons you might expect [1]. The big advance was that it crystallized the need and advantage for a second reference model of the system. With directed testing, the second model was distributed among the test cases in the form of the expected results. Most tests had to be updated when a design or specification change was made. Only one model is affected with a centralized reference model. I will also admit that there is some value in randomization. Being able to explore cases that were not originally considered can bring about some surprises and find bugs that would have been missed by directed tests. Constrained random was also touted as a way to tradeoff the use of highly skilled people for automation, but it turns out, at a high price. The size of server farms and the number of simulator licenses per engineer has gone out of control. Mentor Graphics, one of the principal providers of simulation products based on constrained random, published a study [2] showing that constrained random techniques may produce between 10 times and 100 times more vectors than is actually necessary to obtain desired coverage.

Click here to read more ...