Supported Hardware#
Each prediction runs on a single GPU. Starting in 1.6.0, a container given more than one GPU serves requests from all of them concurrently; refer to Configuration Alternatives.
GPU#
The OpenFold3 NIM is supported to run on the GPU configurations below.
GPU |
GPU Memory |
# of GPUs |
|---|---|---|
NVIDIA H100 80GB HBM3 |
80GB |
1 or more |
NVIDIA H100 NVL |
94GB |
1 or more |
NVIDIA H100 PCIe |
80GB |
1 or more |
NVIDIA A100-SXM4-80GB |
80GB |
1 or more |
NVIDIA A100 80GB PCIe |
80GB |
1 or more |
NVIDIA A100-SXM4-40GB |
40GB |
1 or more |
NVIDIA A100 40GB PCIe |
40GB |
1 or more |
NVIDIA H200 |
141GB |
1 or more |
NVIDIA H200 NVL |
141GB |
1 or more |
NVIDIA B200 |
180GB |
1 or more |
NVIDIA B300 |
288GB |
1 or more |
NVIDIA GB200 |
186GB |
1 or more |
NVIDIA GB300 |
288GB |
1 or more |
NVIDIA GH200 144GB |
144GB |
1 or more |
NVIDIA RTX PRO 6000 Blackwell Server Edition |
96GB |
1 or more |
NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
96GB |
1 or more |
NVIDIA RTX 6000 Ada Generation |
48GB |
1 or more |
NVIDIA L40S |
48GB |
1 or more |
NVIDIA GB10 (DGX Spark) |
128GB |
1 |
Note
The GB10 features 128GB of LPDDR5x unified memory shared between CPU and GPU, providing seamless memory access across the Grace Blackwell Superchip architecture. It is listed as 1 rather than 1 or more because a DGX Spark carries a single GB10 superchip, so there is no second GPU for the NIM to start a worker on.
Storage#
NVIDIA recommends at least 80GB of disk space.
The container is roughly 11GB to download and approximately 40GB once unpacked. Model parameters are not included in the container; they are downloaded on first run into the NIM cache directory, which needs a further 15GB. Refer to Getting Started for how to set the cache location.
CPU and System Memory#
At least 64 GB of RAM
A CPU with at least 8 available cores