Feature Request: ARM64 (Grace CPU) Support for Riva with Whisper Large-v3 Turbo

AI & Data Science Deep Learning (Training & Inference) Riva
1

Environment

  • Hardware: NVIDIA DGX Spark (Grace CPU + Hopper GPU)
  • Architecture: Linux ARM64 (aarch64)
  • Target Riva Version: 2.19.0 or later
  • Model: Whisper Large-v3 Turbo for ASR

Request

I would like to request official ARM64 (aarch64) support for NVIDIA Riva Speech AI, specifically for data center-grade deployments on ARM64 server platforms like DGX Spark with Grace CPU.

Currently, Riva offers ARM64 support primarily for embedded platforms (Jetson series), but I would like to deploy Riva + Whisper Large-v3 Turbo on ARM64 data center servers.

Background

  1. ARM64 adoption in data centers is increasing, with platforms like DGX Spark (Grace + Hopper) becoming more common
  2. The current Riva documentation focuses on x86_64 for data center deployments and ARM64 for embedded use cases
  3. Whisper Large-v3 Turbo’s NGC page does not explicitly mention ARM64 compatibility
  4. When attempting to use Riva on ARM64, some models show “architecture not supported” errors

Use Case

We are running AI workloads on DGX Spark and would like to leverage Riva’s optimized ASR capabilities with Whisper Large-v3 Turbo on the same ARM64 infrastructure, rather than maintaining separate x86_64 nodes for speech processing.

Request Summary

  • Official ARM64 (aarch64) support for Riva Speech AI on data center platforms (not just embedded)
  • ARM64-compatible containers and RMIR models for Whisper Large-v3 Turbo
  • Documentation and guidance for deploying Riva on ARM64 server environments

Thank you for considering this request!

4 Likes