As ephemeral disappears when you stop an instance (presumably, when the system is given the opportunity to relocate your instance to another machine), I've always suspected that ephemeral storage was part of local disks on the virtual host's chassis -- as opposed to a SAN or other kind of network-attached storage. As such, you probably end up with the same unpredictable performance you find in all virtualized resources, since it is very unlikely that Amazon is giving you your own storage.
Before benchmarking ephemeral storage you have to pre-warm it, which might have contributed to your findings. Ephemeral is worlds better than EBS, particularly in outage scenarios; if I could convince everybody on planet Earth to stop using EBS, it would be a noble cause.
Happy to be proven wrong but this is based on a year or so of experience dealing with EBS. You can't see the EBS traffic in your tools (at least that I've been able to find), which complicates things.
It's all kinds of different on VPC instances, therefore I suspect the network interface model -- and possibly EBS connectivity -- is different on those. So, who knows? I'd kill for Amazon to be more forthcoming here so I could understand the infrastructure running my fleet, but, I don't and they aren't.