Estimating divergence times is notorious difficult and the field can be downright rancorous with some being accused of reading tea leaves and chicken entrails - interesting reading for personalities as much as the different scientific approaches. There are several different approaches to trying to estimate a divergence time among species, using calibration points usually anchored by fossil data. Molecular clock methods have problems sometimes producing extremely old dates that are quite hotly debated. In fungi we have a very few fossils (and their placement on the phylogeny is debated).
There are quite a few available methods for reconstructing divergence times including r8s and multidivtime which start with various types of trees and use calibration time points that are typically informed by fossil dates. The simplest approaches assume a molecular clock (rates are same across the tree) and then one only needs to calibrate the number of substitutions (or rate really) to time to determine how phylogenetic tree branch lengths map to time. The BEAST package also does phylogenetic inference and divergence time estimation (and provided the necessary analysis for exoneration of the Tripoli Six) across a sample of trees. BEAST (and MrBayes) use MCMC to sample the space of parameters and tree space to identify phylogenies and evolutionary parameters that explain the data (an alignment of sequences).
A paper from Akerborg and colleagues introduces a new approach that uses MCMC but apply a few twists, using a birth-death model that doesn't assume a molecular clock and employing a hill-climbing algorithm instead of MCMC to find parameter optima. They use a Maximum a posterior (MAP) framework which is more computational efficient than MCMC. They couple the MAP approach with a dynamic-programming approach that separates the estimation of rates (branch length) from the estimation of times (which often require assumption of a molecular clock). While I can't speak with much authority on the MAP approach or yet how well this compares on different datasets, it suggests a different method to tackle these problems. They authors point out one drawback with their approach is it only allows for derivation of point-estimates so statistical confidences like bootstrap support are not easily calculated through this approach. Their software, called PRIME is available here and I will be curious to see how it performs in other peoples' hands.
Akerborg, O., Sennblad, B., Lagergren, J. (2008). Birth-death prior on phylogeny and speed dating. BMC Evolutionary Biology, 8(1), 77. DOI: 10.1186/1471-2148-8-77