VMware has claimed to have achieved “near or better than bare metal performance” with virtualized machine learning workloads.
For the training workloads, VMware used a number of NVIDIA GPUs connected by near-range communications link NVIDIA NVLink.
Its team worked on training the natural language processing workload BERT with the SQuAD dataset, and on training the image segmentation workload Mask R-CNN with the COCO dataset.
What does this mean for users?
Nvidia claim that this solution will give users the benefits of bare metal sever performance while providing all the virtualization-related benefits of VMware including: server consolidation, power savings, virtual machine over-commitment, vMotion, high availability, DRS, central management with vCenter, suspend/resume VMs, and cloning.
VMware’s Uday Kurkure told The Register he expects most high performance computing (HPC) workloads will be virtualized moving forwards, and told the publication that HPC teams are “always running into performance bottlenecks that leaves systems underutilized”.
Kurkere said he anticipates these results to improve resource utilization in a number of different fields, including “investment banking, pharmaceutical research, 3D CAD, and auto manufacturing”.
VMware told The Register it is also investigating how virtualized GPUs perform with even larger AI/ML models, like GPT-3, as well as how these technology be applied to telelcoms workloads running at the edge.
The results were achieved via using Nvidia’s vGPU Manager in vSphere as opposed to the hardware-level partitioning.
VMware’s impressive claims come as reports are circulating of a big money acquisition by chipmaker Broadcom.
Reports have claimed a deal worth $60 billion could be announced on Thursday, the same day as VMware’s financial results. The sources said that it’s likely the deal will take place mostly in cash.