Peter Desantis, vice president of AWS global infrastructure, speaks at AWS re:Invent 2017. (GeekWire Photo / Tom Krazit)

Server virtualization was one of the great breakthroughs in enterprise computing over the last 20 years, allowing people to get far more performance out of a single server by splitting into multiple independent parts. But sometimes you really need the full performance of the server all to yourself, which is why Amazon Web Services is ready to let customers run their workloads on bare-metal servers, the company announced Tuesday night.

The new instances are available as a public preview, and are possible because of design work on custom hardware accelerators accomplished by AWS following the acquisition of Annapurna Labs in 2015, said Peter Desantis, vice president of AWS global infrastructure. AWS used the VMware Cloud on AWS product as a sort-of test run for this technology, and now it’s ready for everybody else.

Unlike most cloud servers, bare-metal servers don’t run typical virtualization software, which allows users to reserve virtual machines running on a server alongside virtual machines from random other AWS customers. Virtual machines are fine for an awful lot of workloads, but they don’t work for everybody, either for performance, compliance, or security reasons.

Bare-metal cloud computing tends to be a specialized product: companies like Packet (which saw this coming in a blog post last week) and Datapipe offer bare-metal computing services, and IBM acquired a bare-metal product along with its Softlayer deal. Microsoft currently doesn’t support bare-metal instances in Microsoft Azure, while Google offers them for specialized GPU (graphics processing unit) workloads on its cloud service.

The new instances were made possible by a project called Nitro, which got going in earnest around 2015 when it decided it needed custom silicon to reach the performance goals it wanted to set for its C5 generation of EC2 instances, the flagship compute instances that made the cloud the cloud. AWS thought about using commercially available chips or FPGAs (field programmable gate arrays) to unlock this performance, but realized it needed a custom product to hit the right level of performance at an acceptable cost, Desantis said.

The acquisition of Annapurna Labs followed soon after that decision, and the C5 instances unveiled earlier this month were the first instances to use the accelerators. The hardware is also part of a redesigned system for EC2 called Nitro, which includes a stripped-down version of the KVM hypervisor that ships standard with a Linux core.

AWS first disclosed the fruits of the Annapurna acquisition last year at re:Invent, when it showed off custom chipsets it had built for its servers using that technology.

Like what you're reading? Subscribe to GeekWire's free newsletters to catch every headline

Job Listings on GeekWork

Find more jobs on GeekWork. Employers, post a job here.