Crusty old VMware is attempting to keep up with the youngsters by acquiring Bitfusion, a startup that claims to enable machine learning on any VM via the magic of network-attached GPUs.
When the deal closes, this capability will be integrated into VMware's vSphere server virtualization platform.
The exact terms of the pact were not disclosed – but Bitfusion has inhaled $8.3m in funding across two rounds since 2015, from investors including Samsung Ventures and Xilinx.
The company is headquartered in Austin, Texas, and peddles something it calls "elastic artificial intelligence", delivered through a software platform called FlexDirect.
Bitfusion said it can enable supercomputing levels of performance for machine learning workloads on clusters of commodity hardware, without any change to the applications themselves.
The upstart has always been close to VMware – its reference architecture, presented at Supercomputing 2018, was based on vSphere and networking hardware from Mellanox.
The traditional approach to machine learning involves GPUs attached directly to individual bare-metal servers. With Bitfusion's platform, network-attached GPUs (NAGs?) scattered across the data centre become a part of a common resource pool, the same way virtualized compute and storage already have done.
Flexible chunks of this GPU resource pool – including fractions of GPUs – can then be used by any virtual machine on the network for either training or inference.
"We view Bitfusion as offering for hardware acceleration what VMware first offered to the compute landscape several years back," said Alex Wang, VMware veep for strategy and corporate development.
FlexDirect supports any type of GPU server and any networking configuration including TCP, RoCE and InfiniBand. The platform also supports other types of accelerators, like FPGAs and ASICs. On the software side, it works with both CUDA and OpenCL frameworks, which means it can deal with cards from either Nvidia or AMD.
Bitfusion said its tech helps maximise GPU utilisation – and with machine learning accelerators often being the most expensive part of the server, this may help reduce the cost of hardware.
It even works in the public cloud. Customers can keep their compute servers running at all times, and only reach for a remote GPU when their workload needs it.
VMware seems to be set on acquiring sexy modern capabilities, rather than building them in-house: in the past six months alone, it announced intent to purchase app-packaging specialist Bitnami and Kubernetes upstart Heptio. ®