Skip to navigation

Boltz-2 Support Matrix

View as MarkdownOpen in Claude

To get started with the Boltz-2 NIM, verify that your machine meets the following hardware and software requirements.

Hardware

Verify that your machine meets the following hardware prerequisites.

Minimum System Hardware Requirements

The following table lists the minimum system hardware requirements.

CPU CoresRAM (GB)NVMe SSD Storage (GB)GPU Count
126480One or more NVIDIA GPUs from the list of supported NVIDIA GPUs

The NVMe SSD storage requirement of 80 GB accounts for the Boltz-2 NIM image size (about 73 GB of disk usage). It also includes additional space for Docker layer caching and temporary files during runtime.

Supported NVIDIA GPUs

The Boltz-2 NIM requires NVIDIA GPUs with at least 48 GB of GPU Memory. While the system can run on any NVIDIA GPU meeting this minimum memory requirement, the following configurations have been specifically tested and optimized with the optimized backend:

GPUGPU Memory (GB)PrecisionGPU Count
NVIDIA A100-SXM4-80GB80BF161-2-4-8
NVIDIA B200180BF161-2-4-8
NVIDIA B300 (NVIDIA-B300-SXM6-AC)288BF161-2-4-8
NVIDIA GB10 (DGX Spark)128BF161
NVIDIA GB200186BF161-2-4-8
NVIDIA GB300 (Grace Blackwell Ultra)288BF161-2-4-8
NVIDIA GH200 144GB144BF161-2-4-8
NVIDIA H100 80GB HBM380BF161-2-4-8
NVIDIA H200141BF161-2-4-8
NVIDIA L40S48BF161-2-4-8
NVIDIA RTX 6000 Ada Generation48BF161-2-4-8
NVIDIA RTX PRO 6000 Blackwell Workstation Edition96BF161-2-4-8

The GB10 uses 128 GB of LPDDR5x unified memory, shared between the CPU and the GPU on the NVIDIA Grace Blackwell Superchip.

For performance and stability, use one of the tested configurations above. Other NVIDIA GPUs with 48 GB or more memory might work, but they have not been tested and might show lower performance or stability.

Model Profile

Release 1.9.0 ships a single untagged model profile. Auto-selection uses this profile on every supported GPU:

ReleaseProfile IDTags
1.9.0af1349b9f5f9a1a6a0404dea36dcc9499bcb25c9adc112b7cc9a93cae41f3262None. Used on all supported GPUs.

Leave NIM_MODEL_PROFILE unset unless you need to pin this profile for reproducibility. Per-GPU profile hashes from 1.8.0 and earlier are not in the 1.9.0 manifest and must not be used. Refer to Select a profile manually.

Software

Install the following software dependencies, then verify that your container runtime can access your local NVIDIA GPUs.

docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi

Example output:

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 590.44.01 Driver Version: 590.44.01 CUDA Version: 13.1 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 NVIDIA H100 80GB HBM3 Off | 00000000:01:00.0 Off | 0 |
| N/A 35C P0 45W / 700W | 0MiB / 81559MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
+-----------------------------------------------------------------------------+

For more information on enumerating multi-GPU systems, refer to the NVIDIA Container Toolkit GPU enumeration topic.

Next Steps

When your hardware and software are in place, you are ready to pull the container and run your first prediction. To get started, refer to the Getting Started guide.