Skip to main content
Version: v2.9.0

Devices supported by HAMi

Applicable version: HAMi v2.9.0.

Last verified: 2026-11-06, against the HAMi source, default device configuration, and this version’s device guides.

The table below summarizes device integration, allocation modes, and multi-card partitioning capabilities for HAMi v2.9.0. See the linked guides for plugin, driver, and runtime requirements.

DeviceModeModelsIntegration statusMemory controlCompute controlMulti-card partitioning
NVIDIA GPUHAMi-coreAllStableMemory quotaTime-slice throttlingYes
NVIDIA GPUMIG (dynamic partitioning)MIG-capable modelsStableHardware profileHardware profileYes (subject to CUDA and driver requirements)
AMD GPUWhole deviceMI300XStableWhole-device allocationWhole-device allocationN/A
AWS NeuronWhole device / NeuronCore allocationInferentia / TrainiumStableNo separate memory quotaNeuronCore allocationNo (multi-device requests allocate all cores of each device)
Cambricon MLUDynamic SMLU370 / 590StableMemory quotaCore quotaNo
Enflame GCUGCUshare (percentage slicing)S60StableShared memory/compute percentageShared memory/compute percentageNo
Ascend NPUtemplate910A, 910B2, 910B3, 910B4-1, 910B4, 310PStableTemplate allocationTemplate allocationNo
Ascend NPUhami-core910A, 910B, 310P, 910C (ARM)StableMemory quotaCompute quotaYes
Ascend NPUWhole device910CStableWhole-device allocationWhole-device allocationN/A
Hygon DCUDynamic vDCU slicingAllStableMemory quotaCompute shareNo
Iluvatar GPUvGPU (gpu-manager)MR-V100, MR-V50, BI-V150, BI-V100StableMemory quotaCore quotaNo
Kunlunxin XPUVXPUP800-OAMStableFixed partition sizesDetermined by partition sizeYes
MetaX GPUsGPUMXC500StableMemory quotaDepends on QoS policyYes (not with sgpu-app-class)
Mthreads GPUsGPUMTT S4000StableDiscrete memory sizesCore quotaNo
Vastai GPUWhole device / Die slicingVA16StableWhole-device or Die allocationWhole-device or Die allocationYes (Die allocation)

Stable means the integration is included in this release.

Capability definitions:

  • Memory control: The memory quota or partition size a container can request. Templates and fixed partitions round requests up to a supported size.
  • Compute control: Devices control compute resources through time slicing, core allocation, hardware partitions, or QoS policies. The linked guides describe the corresponding limits.
  • Multi-card partitioning: A single container can allocate resources across multiple physical devices while using only part of at least one device. Allocating multiple whole devices does not count as multi-card partitioning. See the linked references for allocation granularity and limits. N/A means the mode allocates whole devices only.
CNCFHAMi is a CNCF Incubating project