Backend.AI vs HAMi

Backend.AI is an AI infrastructure platform for building, training, and serving models, as well as running parallel computing jobs. Its container-level GPU virtualization divides physical GPU capacity into fractional allocations so multiple users can share a GPU. The Sokovan scheduler supports multi-node, multi-tenant workloads and policy-based resource management. The platform says its hardware abstraction layer supports more than 12 accelerator types, including products from NVIDIA, AMD, Intel, Rebellions, and FuriosaAI. It can be deployed on-premises, in public clouds, or in a hybrid setup; installation, model deployment, and updates can operate without internet connectivity in air-gapped networks. The browser-based WebUI provides cluster management, resource and session monitoring, and policy-based allocation. Listed cloud environments include AWS, GCP, Azure, and OCI. Storage integrations include VAST, WEKA, Pure Storage, and IBM Storage Scale, with NVIDIA GPUDirect Storage support. Open-source is free for home users installing on their own hardware. Enterprise plans require a quote. Production guidance says compute nodes should be network-isolated with restricted inbound access.

HAMi is open-source middleware for sharing and scheduling GPUs and other accelerators among AI workloads on Kubernetes. Workloads can request GPU memory and core limits through Kubernetes resource controls. Scheduling options include binpack, spread and numa-first, with binpack and spread available at node or GPU scope. The supported-device list spans NVIDIA, AMD, AWS Neuron, Huawei Ascend, Cambricon, Enflame, Hygon, Iluvatar, Kunlunxin, MetaX, Moore Threads, Vastai and Biren; isolation and partitioning capabilities vary by device. HAMi works with Kubernetes APIs, DRA and CDI, and documents scheduling integrations including Volcano, Kueue, Koordinator and NVIDIA KAI Scheduler. Its node metrics endpoint exposes device memory and core utilization for Prometheus. Classic HAMi requires Kubernetes v1.23 or later. HAMi-DRA requires Kubernetes v1.34 or later, the DRA Consumable Capacity feature gate, CDI and NVIDIA driver 440 or later. NVIDIA limits use user-space library interception, which does not cover applications that bypass CUDA. Enforcement is described as soft and best-effort; the FAQ recommends MIG when hardware-enforced isolation is required.