Skip to content

Supported Accelerators ​

Document Information

Version: v1.0 Updated: 2026-07-13

How to Read This List ​

The models below are within the current AGIOne accelerator-management scope. Inclusion means that the accelerator model can be evaluated for On-Prem onboarding; it does not guarantee that every driver, runtime, image, inference engine, model, quantization method, or multi-node topology is compatible.

Before delivery, freeze and validate the complete combination:

Accelerator -> Operating system -> Driver and runtime -> Container image -> Inference engine -> Model -> Deployment topology

Accelerator List ​

VendorArchitecture / SeriesModels
NVIDIAHopperH800, H200, H100, H20
NVIDIAAmpereA100, A800, A40, A30, A10, RTX A6000, RTX A5000, RTX A4000, RTX A2000, RTX 3090, RTX 3060
NVIDIAAdaL40, L40S, L20, L20S, L4, L2, RTX 6000, RTX 5000, RTX 4500, RTX 4000, RTX 2000, RTX 4090, RTX 4090D
Huawei AscendAscend 910Ascend 910B, Ascend 910C
EnflameEnflameS60
BirenBiren106
HygonBWBW200

Required Project Validation ​

CheckWhat to Confirm
Hardware inventoryExact model, card count, device memory, server architecture, and topology
Software stackOperating system, kernel, driver, CUDA/CANN/vendor runtime, and device plugin
Runtime assetsContainer runtime, image, inference engine, model weights, and storage path
Model fitRequired memory, precision or quantization, parallelism, context length, and expected throughput
Cluster readinessNode status, device registration, scheduling labels, storage, registry, network, and time synchronization
AcceptanceDeployment success, model-call correctness, monitoring visibility, stability, and target performance

For accelerators not listed above, perform an adaptation assessment before committing to onboarding or model deployment.

Separate Cloud and Accelerator Support

Huawei Ascend accelerator onboarding is an AI Infra On-Prem compatibility topic. It does not mean that Huawei Cloud access is supported. Huawei Cloud access is currently temporarily unsupported.