NVIDIA H800, H20, and H20A confidential-computing support
NVIDIA’s live matrix covers specific H800, H20, and H20A variants. This is a regional-product support page, not a claim that every Hopper derivative or provider instance is confidential.
Direct answer
NVIDIA currently lists H800 PCIe, H800 NVL, named HGX H800 systems, HGX H20 8-GPU 141 GB, and HGX H20A 8-GPU 96 GB in its confidential-computing support materials. Single-GPU passthrough is broadly listed; Protected PCIe multi-GPU support is release-row specific, and no named cloud CC product was confirmed.Why these models are grouped
H800, H20, and H20A are regional/export-market Hopper products with similar confidential-computing architecture and overlapping deployment requirements. Grouping avoids repeating the same CPU TEE, passthrough, attestation, and operational constraints while preserving stable sections for every verified product.
H800 PCIe
The Secure AI Compatibility Matrix names H800 PCIe for single-GPU passthrough confidential mode. Verify the exact board/VBIOS/firmware row and supported driver rather than assuming every H800 PCIe listing qualifies.MATRIX
H800 NVL
H800 NVL also appears as a named single-GPU passthrough product. First-party cloud documentation does not expose a named H800 NVL confidential-computing service, so its current availability is vendor-supported hardware rather than public-cloud GA.MATRIX
HGX H800 8-GPU
The live matrix lists air-cooled and partner-cooled HGX H800 8-GPU variants. R595 explicitly names Protected PCIe support for the air-cooled HGX H800 configuration. Treat cooling/system rows as material and check the current matrix before promising multi-GPU mode on a partner platform.R595MATRIX
HGX H20 8-GPU 141 GB
HGX H20 8-GPU 141 GB appears in the live support matrix for confidential computing. The live matrix contains a Protected PCIe entry that the R595 feature summary does not name. That difference is a reason to bind a procurement statement to the exact live row, firmware and driver—not to advertise general H20 multi-GPU support from this page alone.R595MATRIX
HGX H20A 8-GPU 96 GB
NVIDIA R595 and the live matrix name HGX H20A for confidential operation, including Protected PCIe multi-GPU support on the documented 8-GPU system. The same Hopper peer-link caveat applies: GPU-to-GPU NVLink/NVSwitch traffic is not encrypted.R595MATRIX
| Model/family | Single-GPU CC | Protected PCIe multi-GPU publication position |
|---|---|---|
| H800 PCIe | Listed | Not generalized; check current row |
| H800 NVL | Listed | Not generalized; check current row |
| HGX H800 8-GPU | Listed | R595 names air-cooled variant |
| HGX H20 8-GPU 141 GB | Listed | Live matrix has a row; R595 summary differs—verify exact release |
| HGX H20A 8-GPU 96 GB | Listed | R595 names support |
Shared Hopper security and deployment constraints
- Requires a supported Intel TDX or AMD SEV-SNP/SEV host path as documented for the platform.
- Requires CC-On and successful CPU/GPU attestation before secret or data release.
- Uses passthrough; current Confidential Containers support excludes MIG and vGPU.
- Hopper Protected PCIe does not encrypt GPU peer traffic over NVLink/NVSwitch.
- Transfer-heavy workloads and pinned-memory assumptions require performance and compatibility testing.
- No named public-cloud CC offering was confirmed for these regional variants.
Frequently asked questions
Does NVIDIA support confidential computing on H800?
Yes, on named H800 PCIe, H800 NVL, and HGX H800 rows. Multi-GPU support is system/release specific, so verify the exact compatibility row.
Is H20 multi-GPU confidential mode supported?
The live matrix includes an H20 Protected PCIe row, while the R595 summary omits it. Treat this as an exact-version verification requirement, not a broad promise.
Which cloud offers confidential H800, H20, or H20A?
Official cloud documentation does not identify a provider-level confidential-computing product for these GPUs.
Sources
- R595
- MATRIX
- PLATFORMS
- DEPLOY