Support Matrix for NVIDIA NIM for Image OCR (NeMo Retriever OCR)#

This documentation describes the software and hardware that NVIDIA NIM for Image OCR (NeMo Retriever OCR) supports.

System Requirements#

Use a supported NVIDIA GPU system with a current NVIDIA driver and NVIDIA Container Toolkit. For validated data center systems, see the NVIDIA Certified Systems Catalog.

Model#

Publisher

Model

Model ID

Variants

Model Card

NVIDIA

NeMo Retriever OCR v2

nvidia/nemotron-ocr-v2

English, multilingual

Model card

The multilingual variant is the default. Set NIM_ENGINE_MODEL_VARIANT=english to serve the English variant.

Supported Hardware and Memory Footprint#

The NIM supports the following validated GPU SKUs & Precision. Where available, the approximate GPU memory sizes are the startup measurements from the latency configuration after forward buffers are allocated. Runtime GPU memory use can increase with larger batches, additional OCR engines, CUDA graph capture shapes, and input-dependent working buffers.

GPU SKU

GPU

English approximate GPU memory size (GiB)

Multilingual approximate GPU memory size (GiB)

Precision

NVIDIA-A100-SXM4-80GB

NVIDIA A100 SXM4 80GB

1.79

2.88

FP16

NVIDIA-A10G

NVIDIA A10G

1.63

2.72

FP16

NVIDIA-B200

NVIDIA B200

2.17

3.26

FP16

NVIDIA-GB10

NVIDIA GB10 Grace Blackwell Superchip, such as NVIDIA DGX Spark systems

N/A

N/A

FP16

NVIDIA-GB200

NVIDIA GB200

2.23

3.33

FP16

NVIDIA-H100-80GB-HBM3

NVIDIA H100 80GB HBM3

2.07

3.17

FP16

NVIDIA-H100-NVL

NVIDIA H100 NVL

2.07

3.17

FP16

NVIDIA-H200

NVIDIA H200

2.07

3.17

FP16

NVIDIA-L4

NVIDIA L4

1.57

2.66

FP16

NVIDIA-L40S

NVIDIA L40S

1.80

2.89

FP16

NVIDIA-RTX-PRO-6000-Blackwell-Server-Edition

NVIDIA RTX PRO 6000 Blackwell Server Edition

1.94

3.03

FP16

Note

GB10 uses unified system memory. GPU memory usage is reported as N/A because the usual VRAM measurements are not accurate for this architecture. For more information, refer to Guidance for Reporting Memory Resources with Unified Memory Architecture.

Throughput mode defaults to two OCR engines and a batch size of 16, which requires more GPU memory than latency mode.

Software#

NVIDIA Driver#

Install a current NVIDIA driver that supports the CUDA version packaged with the NIM. If the container fails to start, verify that the driver and NVIDIA Container Toolkit are current.

NVIDIA Container Toolkit#

Your Docker environment must support NVIDIA GPUs. For installation instructions, see the NVIDIA Container Toolkit documentation.