HAMi
project-hami.io
- Free plan
- Linux
- Self-hosted
What HAMi is
HAMi is open-source middleware for sharing and scheduling GPU resources across AI workloads on Kubernetes. Workloads can request GPU memory and core limits through Kubernetes resource limits, while schedulers can use binpack, spread or numa-first policies. Binpack and spread can be applied at node or GPU scope. The project lists accelerators from NVIDIA, AMD, AWS Neuron, Huawei Ascend and other vendors, though isolation and partitioning capabilities vary by device. HAMi integrates with Kubernetes APIs, DRA and CDI, and documents links to 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 v1.34 or later, plus the DRA Consumable Capacity feature gate, CDI and NVIDIA driver 440 or later. NVIDIA limits rely on user-space library interception, so applications that bypass CUDA are not covered. The FAQ describes memory and compute enforcement as soft and best-effort.
Who it suits
HAMi suits teams running AI, machine-learning or HPC workloads on Kubernetes that need to share heterogeneous accelerators. Teams requiring hardware-enforced isolation for compliance or SLAs should account for its stated limits and the recommendation to use MIG.
Good to know, and what to watch
- Free, open-source GPU virtualization middleware.
- Workloads can request GPU memory and core quotas.
- Supports multiple scheduling policies and integrations.
- Exposes utilization metrics for Prometheus scraping.
- HAMi-DRA requires Kubernetes v1.34 or later.
- Some devices lack isolation or partitioning capabilities.
- NVIDIA applications bypassing CUDA are not covered.
- Priority field supports only two levels.
Plans
Every plan and priceHAMi
Free plan
Getting started
- Prepare a Linux self-hosted Kubernetes cluster
- Use Kubernetes v1.23 or later for classic HAMi
- Install Helm and kubectl, and obtain installation permissions
- Install HAMi with Helm
- For HAMi-DRA, use Kubernetes v1.34 or later with the required feature gate, CDI, and NVIDIA driver 440 or later
- Configure GPU memory and core limits through Kubernetes resources
Limits to check
HAMi describes vGPU memory and compute enforcement as soft and best-effort; applications bypassing the CUDA library are outside NVIDIA interception. Supported-device capabilities vary, and some listed devices lack memory isolation, core isolation, or multi-card partitioning.
The verdict
HAMi offers Kubernetes-based accelerator sharing, resource controls and scheduling across a broad device list. Its best-effort isolation and device-specific capability limits matter for teams with strict isolation requirements.
Questions people ask
What does HAMi cost?
The HAMi plan is listed as free. It is open-source GPU virtualization middleware.
Which environments does HAMi support?
HAMi runs on Linux in a self-hosted deployment and supports Kubernetes. Classic HAMi requires Kubernetes v1.23 or later.
Which accelerators are supported?
The project lists NVIDIA, AMD, AWS Neuron, Huawei Ascend, Cambricon, Enflame, Hygon, Iluvatar, Kunlunxin, MetaX, Moore Threads, Vastai, and Biren devices.
How are GPUs scheduled?
Scheduling policies include binpack, spread, and numa-first. Binpack and spread can be applied at node or GPU scope.
Does HAMi provide monitoring?
Yes. A node metrics endpoint exposes real-time device memory and core utilization that Prometheus can scrape.
When is MIG recommended?
HAMi recommends MIG when hardware-enforced isolation is required for compliance or service-level agreements.
What the maker publishes
- Accelerators
- The project lists support for NVIDIA, AWS Neuron, Huawei Ascend, Cambricon, Enflame, Hygon, Iluvatar, Kunlunxin, MetaX, Moore Threads, Vastai, AMD and Biren accelerators. source
- Community support
- The project links to community support through Discord and Slack (#hami-dev). source
- Deployment
- The quick start installs HAMi with Helm and requires Kubernetes, Helm, kubectl, and installation permissions. source
- Deployment requirements
- Classic HAMi supports Kubernetes v1.23 or later; HAMi-DRA requires Kubernetes v1.34 or later with the DRA Consumable Capacity feature gate enabled, CDI and NVIDIA driver 440 or later. source
- Device coverage
- The v2.10.0 supported-device matrix lists NVIDIA, Cambricon, Hygon, Huawei Ascend, Iluvatar, Mthreads, MetaX, Enflame, Kunlunxin, Vastai, AMD, AWS Neuron, and Biren devices as stable. source
- Ecosystem integrations
- The project documents integrations with Volcano, Kueue, Koordinator, and NVIDIA KAI Scheduler. source
- Integrations
- The site lists Kubernetes, Volcano, Kueue, Koordinator and KAI Scheduler in its Kubernetes scheduling ecosystem. source
- Isolation limits
- The FAQ characterizes HAMi vGPU memory and compute enforcement as soft and best-effort, and recommends MIG when hardware-enforced isolation is required for compliance or SLAs. source
- Isolation mechanism
- For NVIDIA devices, HAMi enforces limits through user-space library interception; the FAQ says applications that bypass the CUDA library are not covered. source
- Kubernetes integration
- HAMi works with Kubernetes APIs, DRA, and CDI. source
- Limitations
- The supported-device matrix marks memory isolation, core isolation, and multi-card partitioning as unavailable for some listed devices. source
- Monitoring
- HAMi exposes real-time device memory and core utilization metrics at a node metrics endpoint that can be scraped by Prometheus. source
- Monitoring
- HAMi provides allocation counts and spread plus real-time GPU memory and core utilization visibility. source
- Product
- HAMi is open-source GPU virtualization middleware that enables sharing, isolation and scheduling of heterogeneous accelerators for AI workloads on Kubernetes. source
- Project status
- HAMi was accepted as a CNCF Incubating project on July 2, 2026. source
- Project status
- HAMi is a CNCF Incubating project. source
- Purpose
- HAMi is open-source, cloud-native GPU virtualization middleware for sharing, isolating, and scheduling heterogeneous accelerators on Kubernetes. source
- Resource controls
- HAMi supports GPU memory and compute quotas, with hard runtime isolation for supported devices. source
- Resource slicing
- HAMi lets workloads request GPU memory and core limits using Kubernetes resource limits such as `nvidia.com/gpumem` and `nvidia.com/gpucores`. source
- Scheduling
- Scheduling policies include binpack, spread and numa-first; binpack and spread can also be selected at node or GPU scope. source
- Scheduling
- HAMi offers binpack, spread, and topology-aware scheduling policies. source
- Scheduling limit
- HAMi's built-in priority field supports two levels; the FAQ recommends integrating Volcano for multi-level queue priorities. source
- Support
- The site links to documentation, tutorials, Discord and the `#hami-dev` Slack channel for community resources. source
- Workloads
- The site identifies LLM, machine-learning, and HPC workloads as use cases. source
Checked against the maker's pages on 4 Oct 2026.