Support Matrix#

Important

The Riva SDK release only supports embedded (L4T) platforms. For x86 data center deployments, refer to Riva NIM documentation.

These support matrices list the supported hardware and software requirements of Riva.

Riva 2.24.0#

Hardware Compatibility#

Embedded (aarch64)

Operating System

Riva server requires L4T aarch64.

GPU Model

Deployment Platforms:

Jetson SDK version

  • JetPack 7.0

Software Compatibility#

Embedded (aarch64)

Container

Docker

  • Docker >= 28.0 (pre-installed with JetPack 7.0)

NeMo

Model Information#

Service

Model

Batch Size

Inference Mode

Unified Memory(GB)

ASR

parakeet-1-1b-ctc

1

streaming
silero_diarizer

4.3
6.8

ASR

parakeet-1-1b-rnnt-unified-ml-cs

1

streaming

12

TTS

magpie

1

streaming

8.5

Skills Clients#

Riva Speech AI Skills clients do not require a local GPU and have minimal hardware requirements. Refer to Clients in a New Programming Language for creating client bindings in your programming language. The Python section describes the prebuilt Python bindings included in the Riva Quick Start package, which are also based on gRPC.

Riva 2.19.0#

Hardware Compatibility#

Data Center (x86)

Embedded (aarch64)

Operating System

Riva server requires Linux x86_64 or Linux aarch64.

Riva server requires L4T aarch64.

GPU Model

Preferred Deployment Platforms:

Note

Riva is supported on any NVIDIA Turing or later GPU for development purposes. Care must be taken to not exceed the memory available when selecting models to deploy. 16+ GB VRAM is recommended.

Deployment Platforms:

Jetson SDK version

  • JetPack 6.0

Software Compatibility#

Data Center (x86)

Embedded (aarch64)

Container

Docker

Note

For DGX users, refer to Preparing to use NVIDIA Containers.

MIG (Multi Instance GPU)

  • Deployment on MIG devices is possible on supported GPUs, provided enough resources are available as per the selected models.

Note

For further information, refer to MIG User Guide.

NeMo

Helm

NVIDIA Driver

Note

For earlier driver versions, refer to the NVIDIA Driver section in Deep Learning Frameworks Support Matrix.

Model Information#

Service

Model

Batch Size

Inference Mode

CPU Memory (GB)

GPU Memory (GB)

Supported on Embedded devices

ASR

conformer-ctc

64

offline,
streaming,
streaming-throughput

2,
2,
2

5.8,
3.6,
4.2

✅

ASR

parakeet-0-6b-ctc

64

offline,
streaming,
streaming-throughput

3,
3,
3

5.8,
4,
5

✅

ASR

parakeet-1-1b-ctc

64

offline,
streaming,
streaming-throughput

5,
5,
5

6.7,
5,
5.9

✅

ASR

parakeet-1-1b-rnnt

64

offline

3

8.5

❌

ASR

canary-1b

8

offline

6.5

13.4

❌

ASR

canary-0.6b-turbo

8

offline

5.3

12.2

❌

ASR

whisper-large-v3

8

offline

4.3

12.5

❌

ASR

whisper-large-v3-turbo

8

offline

3.8

11.3

❌

NMT

megatron-1b

8

offline

9

✅

TTS

fastpitch-hifigan

8

offline,
streaming

2

✅

TTS

magpie

8

offline,
streaming

5.8

❌

Skills Clients#

Riva Speech AI Skills clients do not require a local GPU and have minimal hardware requirements. Refer to Clients in a New Programming Language for creating client bindings in your programming language. The Python section describes the prebuilt Python bindings included in the Riva Quick Start package, which are also based on gRPC.