[flexiondotorg]

Passing through a GPU to a kernel in a VM requires fiddling with SR-IOV and IOMMU, which also requires your hardware is compatible. Often hardware is not compatible, either by design because these features are artificially reserved for "enterprise" models or due to accidental IOMMU group conflicts on lower priced motherboards.

As containers share the host kernel, these issues do not arise and hardware acceleration is trivial.