Virtualizing Nvidia Hgx B200 Gpus with Open Source

Isn't SR-IOV a thing with these big GPUs? Or, is it that you're not concerned with fractional granularity?

Would it be possible to implement "virtual memory" for a GPU this way? Let's say you have GPUs at 30% utilization, but memory limited. Could you run 2 workloads by offloading the GPU memory when not in use?

Thanks! I haven't looked deeply into slicing up a single GPU. My understanding is that vGPU (which we briefly mention in the post) can partition memory but time-shares compute, while MIG is the only mechanism that provides partitioning of both SMs and memory bandwidth within a single GPU.

Last I checked MIG was the only one that made hard promises about especially memory bandwidth; as long as your memory access patterns aren't secret and you have enough trust in the other guests not being highly unfriendly with their cache usage behavior, you should be able to get away with much less strict isolation. Think docker vs. VMs-with-dedicated-cores.

But I thought MIG did do the job of chopping a GPU that's too big for most individual users into something that behaves very close to a literal array of smaller GPUs stuffed into the same PCIe card form factor? Think how a Tesla K80 was pretty much just two GK210 "GPUs" on a PLX "PCIe switch" which connects them to each other and to the host. Obviously trivial to give one to each of two VMs (at least if the PLX didn't interfere with IOMMU separation or such.... for mere performance isolation it certainly sufficed (once you block a heavy user from power budget throttling the sibling, at least).

Fabric Manager itself is not open source. It's NVIDIA-provided software, and today it's required to bring up and manage the NVLink/NVSwitch fabric on HGX systems. What we meant by "open" is that everything around it - the hypervisor, our control plane logic, partition selection, host configuration, etc. - is implemented in the open and available in our repos. You're right that this isn't a fully open GPU stack.

On isolation: in Shared NVSwitch Multitenancy mode, isolation is enforced at multiple layers. Fabric Manager programs the NVSwitch routing tables so GPUs in different partitions cannot exchange NVLink traffic, and each VM receives exclusive ownership of its assigned GPUs via VFIO passthrough. Large providers apply additional hardening and operational controls beyond what we describe here. We're not claiming this is equivalent to AWS's internal threat model, but it does rely on NVIDIA's documented isolation mechanisms.

Thanks for the comment! You're right that a lot of the mechanics apply more generally. On point (3) specifically: we handle this by allocating at the IOMMU-group level rather than individual devices. Our allocator selects an IOMMU group and passes through all devices in that group (e.g., GPU video + audio), which avoids the partial-passthrough wonkiness you mentioned. For reference: https://github.com/ubicloud/ubicloud/blob/main/scheduling/al...

We haven't looked deeply at inter-machine communication yet. NVLink/NVSwitch (which this post focuses on) are intra-node, so InfiniBand is mostly orthogonal I think and comes down to NIC passthrough, NUMA/PCIe placement, and validating RDMA inside the VM.

We didn't focus on vGPU and largely avoided it on purpose. Instead, we focused on whole-GPU and NVSwitch-partitioned passthrough (Shared NVSwitch Multitenancy Mode), which is a better fit for the workloads we care about.

Base Address Register

this term can be used at a couple different points (including mappings from physical addresses to physical hardware in the memory network), but a PCI BAR is a register in the configuration space that tells the card what PCI host addresses map to internal memory regions in the card. one BAR per region.

A couple open relevant issues here:

https://github.com/amd/MxGPU-Virtualization/issues/6

https://github.com/amd/MxGPU-Virtualization/issues/16