DynamiaDynamia
GPU COMPATIBILITY

Heterogeneous GPU compatibility

Explore supported NVIDIA, AMD, AWS and Chinese accelerators, including device families, resource-sharing capabilities and deployment requirements.

Supported GPUs and AI accelerators

NVIDIA

Capabilities and deployment requirements

Supports memory and compute sharing, isolation, multi-GPU virtualization and Binpack / Spread scheduling. Requires drivers compatible with the GPU and CUDA, NVIDIA Container Toolkit and the HAMi device plugin. Disable GPU Operator’s native device plugin when using it.

Read integration guide

AMD

Capabilities and deployment requirements

Supports concurrent GPU sharing, memory allocation in MiB and proportional compute-unit (CU) allocation. Requires AMD drivers, ROCm and the HAMi AMD device plugin. Shared workloads need glibc 2.34 or later; disable AMD GPU Operator’s native device plugin.

Read integration guide

AWS

Capabilities and deployment requirements

Supports NeuronCore-based resource allocation, device sharing and topology-aware multi-device scheduling for training and inference. Deploy on Inf / Trn EC2 instances with Neuron drivers and runtime, Neuron Device Plugin and HAMi.

Read integration guide

Huawei Ascend

Capabilities and deployment requirements

Template-based hardware partitioning or software sharing. Memory and compute capabilities depend on the mode; software sharing requires compatible ARM hardware, drivers and the HAMi scheduler.

Read integration guide

Hygon

Capabilities and deployment requirements

Memory and compute sharing; memory isolation is model-dependent. Requires the vendor device plugin and compatible DTK/drivers. Check the single-container vDCU limit.

Read integration guide

Cambricon

Capabilities and deployment requirements

Memory and compute sharing. The MLU370 guide requires Dynamic SMLU and matching drivers/CNToolkit; verify the applicable versions for other models.

Read integration guide

MetaX

Capabilities and deployment requirements

Memory and compute sharing via sGPU. Enable vendor virtualization and install the device plugin for the specific model.

Read integration guide

Iluvatar CoreX

Capabilities and deployment requirements

Memory and compute sharing; the guide specifies 256 MiB memory and 1% compute granularity. Requires GPU Manager and compatible drivers; verify device limits per container.

Read integration guide

Moore Threads

Capabilities and deployment requirements

Memory and compute sharing. Partition granularity and isolation depend on the vendor driver and deployment guide.

Read integration guide

Kunlunxin

Capabilities and deployment requirements

Whole-device scheduling and a VXPU path. Memory sharing does not imply independent compute partitioning. Whole-device scheduling is topology-aware; select the deployment mode explicitly.

Read integration guide

Enflame

Capabilities and deployment requirements

Memory and compute sharing. Allocate resources using supported vGCU/device-plugin configurations; verify driver and device mode.

Read integration guide

Vastai

Capabilities and deployment requirements

Listed as schedulable in the public matrix, without memory or compute partitioning. Plan for whole-device allocation and confirm any sharing requirements separately.

Read integration guide

Biren

Capabilities and deployment requirements

Listed as schedulable in the public matrix, without memory or compute partitioning. Confirm the model, plugin and delivery scope for the target environment.

Read integration guide

T-Head

Capabilities and deployment requirements

Supports fine-grained memory and compute isolation within unified compute management. Integrates through the Dynamia PPU device plugin with Zhenwu drivers, runtime and Dynamia product components.

Dynamia product documentation

Find the right compute solution

Tell us about your GPU models, cluster environment and workloads. Together, we can find the right approach to resource sharing, scheduling and management.

Image viewer