the benchmarks with the direct method calls use the injected beans directly. that means those benchmarks show the raw performance. the benchmarks with the el-resolver calls use the jsf-api for resolving the beans and therefore they show the performance of the el-resolver implementations of both cdi implementations.
this benchmark uses the weld version which is bundled with glassfish 3.1:
this benchmark uses the (currently) latest snapshot of weld (downloaded from the snapshot repository):
owb has a special proxy for request scoped beans - that's the reason why there is such a huge difference.