Support Matrix

Each design request runs on a single GPU.

GPU

The RFD4-Proteina NIM has been validated on:

GPU	GPU Memory	# of GPUs
NVIDIA H100 80GB HBM3	80GB	1, 4, 8
NVIDIA H100 NVL	94GB	1, 4, 8
NVIDIA H100 PCIe	80GB	1, 4, 8
NVIDIA H200	141GB	1, 4, 8
NVIDIA H200 NVL	141GB	1, 4, 8
NVIDIA A100-SXM4-80GB	80GB	1, 4, 8
NVIDIA A100 80GB PCIe	80GB	1, 4, 8
NVIDIA A100-SXM4-40GB	40GB	1, 4, 8
NVIDIA A100 40GB PCIe	40GB	1, 4, 8
NVIDIA GH200 144GB HBM3e	144GB	1, 4, 8
NVIDIA B200	180GB	1, 4, 8
NVIDIA B300	288GB	1, 4, 8
NVIDIA GB200	186GB	1, 4, 8
NVIDIA GB300	288GB	1, 4, 8
NVIDIA RTX PRO 6000 Blackwell Server Edition	96GB	1, 4, 8
NVIDIA RTX PRO 6000 Blackwell Workstation Edition	96GB	1, 4, 8
NVIDIA RTX 6000 Ada Generation	48GB	1, 4, 8
NVIDIA L40S	48GB	1, 4, 8
NVIDIA GB10 (DGX Spark)	128GB	1

Note

Each design request runs on a single GPU; “# of GPUs” in the preceding table is the common server configurations this NIM has been deployed on, so a worker pool can serve several requests at once across more than one GPU. Refer to Configuration Alternatives for running with more than one GPU. The GB10 (DGX Spark) carries a single GB10 superchip, so it is listed as 1.

This NIM ships with no refolding oracle built in (refer to Oracles); a mounted oracle that calls an independently hosted structure-prediction NIM adds no GPU memory cost of its own here, since the model it calls lives entirely in that other service.

Minimum System Hardware Requirements

CPU Cores	RAM (GB)	NVMe SSD Storage (GB)	GPU Count
12	16	60	1

The container downloads model checkpoints into the NIM cache directory on first run; provide a NIM cache location with sufficient free space. Refer to Getting Started for how to set the cache location.

Operating System / Architecture

The container image publishes both linux/amd64 and linux/arm64 manifests, and pulls and serves on both: amd64/x86_64 hosts (H100, H200, A100, L40S, RTX 6000 Ada, RTX PRO 6000 Blackwell) and arm64/aarch64 Grace hosts (GB200, GB300, GB10/DGX Spark), all listed in the GPU table above. Any Docker-supported Linux distribution on either architecture is expected to work; the amd64 image is built FROM the nvcr.io/nvidia/pytorch:26.05-py3 base image, and no distro-specific behavior is introduced on top of it.

Supported Tasks and Input Dialects

Task	Requires an input structure	Requires a ligand
generation	No	No
motif_scaffolding	Yes	No
enzyme_design	Yes	No
small_molecule_binder	Yes	Yes
antibody_design	Yes	No
partial_diffusion	Yes	No
inverse_folding	Yes	No

Two input dialects are supported for every task that requires a structure:

• annotated_cif: a single self-contained annotated mmCIF file.

• condition_spec: a structured condition specification (YAML/JSON) plus the structure file(s) it references.

Refer to Example Requests for both dialects in use, and /v1/tasks (refer to Getting Started) for what a running deployment currently serves.