If it's RSEM from the archive version of TCGA then there is a field in there which is similar to TPM.
See this entry:
If someone ran RSEM more recently and you got that data from them then there is probably something like TPM in there somewhere.
Since TPM is normalized for gene length it is fair to compare two different genes from the same sample.
BUT there are some caveats. Different genes get quantified differently, i.e. due to different GC content or length so small differences should be considered "ehhh" even if they are statistically significant because the bias is expected to be systematic. Huge differences, like a 10X fold change should be reliable.