Alfa-Bank Brings InfiniBand and GPUDirect into Kubernetes: GPUaaS in 3 Months

⚡ The gist in 5 seconds

• Alfa-Bank has put into production a GPUaaS on Kubernetes with InfiniBand, RDMA and GPUDirect — three months from zero to a working stack.
• Teams are separated by InfiniBand partitions (Partition Key), the OS is provisioned with Ansible, quotas and resources are managed by Argo CD.
• Everything is built on open source with no licensed solutions; GPU Operator is configured via MIG and Time Slicing.

🔍 What happened

An engineer going by the nickname DevEvgenii, in Alfa-Bank's blog on Habr, broke down how the bank built and put into production a GPUaaS — a shared GPU pool for AI teams, on-prem and in the cloud — in three months. The task wasn't just "installing GPUs" but getting the InfiniBand + RDMA + GPUDirect combo working inside Kubernetes. It's usually set up on bare metal, and, according to the author, there are no ready-made recipes for an orchestrator in the Russian-speaking community — the article on Habr.

The mechanics, step by step. DOCA-Networking packages are installed on the host — including doca-ofed, roce and host-basic — or the full DOCA-ALL, plus nvidia-driver, nvidia-fabricmanager and nvidia-utils. Then the GPU Operator with MIG and Time Slicing. Teams are separated by InfiniBand partitions: access to someone else's GPU memory and interfaces is blocked at the Partition Key level. The routine was eliminated via GitOps — the OS is provisioned with Ansible, while quotas and resources are handled by Argo CD.

🧠 Our take

The reason this write-up exists at all is economics. GPUs are expensive, and they start sitting idle precisely when each project buys its own and waits months for delivery. The author states it plainly: the platform solves the problem of expensive hardware sitting idle. To me, that's the main argument, not the three-month timeline — the timeline actually looks normal for an infrastructure team that already knows Kubernetes.

What really changes here is multi-tenancy. Usually GPUs in a cluster are handed out "as is," and your neighbor on the same hardware sees more than you'd like. Isolation of GPU memory via MIG and Partition Key is the hardest part to verify, and it's the one most often postponed. Our assessment: the value of the article isn't in "we did it," but in listing the specific packages and tools you can replicate yourself.

Who wins is clear: AI teams that get capacity from a pool instead of through procurement. Who loses — honestly, I'm not sure anyone does. Except perhaps those who will keep building bare-metal setups for every project. In my experience, such stacks survive exactly until the first person leaves and takes the knowledge of how everything was assembled with them. GitOps is precisely the insurance against that.

⚠️ What we don't know

The article contains not a single number about the hardware: how many GPUs, which models, how many nodes, what RDMA throughput in benchmarks, or what it all cost. The only figure is the timeline — three months. The claim that "there's little expertise on the market" can't be verified: it's the author's word, not a study. There's a single source — the article itself — and it should be treated as a good case study, not a benchmark.

Prediction: Within the quarter, at least one more public case of GPUaaS on Kubernetes with GPUDirect RDMA from a Russian company will emerge — if GPU scarcity and idle hardware remain the market's main pain point.

Source: Habr / Alfa-Bank