But maybe that not surprising since Hypertable was designed by Zvents search architect Doug Judd for speed. He rejected Java (used by HBase, the Hadoop-project Bigable effort) in favor of C++ in order to get the performance as high as possible.
With a small test inserting about 28M rows of data from the AOL search dataset, they achieved a per-node write rate of approximately 7mb/sec. Iteration over the data once loaded was also quite fast, at nearly 1M cells/second.
The question is how the system will scale up to much larger amounts of data. But the early perf numbers are encouraging. Doug and co will also need to get the word out about Hypertable and get a developer community going around this project if it's going to achieve its full potential.
Hypertable can run on top of either HDFS or KFS. Zvents CEO Ethan Stock told me they will be releasing it under GPL 2.1 on Jan 31th.