Frank Fu's Blog

NavTalk’s digital human lip-sync and real-time audio/video capabilities now fully support deployment and operation on Linux servers equipped with NVIDIA RTX 5090. The full stack—from drivers and frameworks to the inference engine—has been adapted and validated for the latest Blackwell-based compute and NVIDIA driver/runtime stack, delivering stable, high-performance real-time digital human experiences on current hardware.

This document describes NavTalk’s official support for 5090 on Linux in three areas: technology stack, adaptation work, and product value, for evaluation and selection reference.

1. Why 5090 and Linux Matter

▪️ Compute upgrade: RTX 5090 is built on the Blackwell architecture, with significantly higher VRAM and compute, suited for real-time high-resolution lip-sync and multi-session concurrency.

▪️ Linux-first: Most production and cloud environments run Linux; NavTalk offers a full set of service modes on Linux (including real-time lip-sync, video lip-sync, and other APIs), making integration and scaling straightforward.

▪️ Long-term alignment: Adaptation is complete for the latest NVIDIA drivers and AI runtimes (e.g. CUDA 12.8, PyTorch 2.7), keeping NavTalk aligned with the official software stack for the foreseeable future and reducing upgrade friction.

Thus, “deployable, operable, and scalable” on 5090 Linux is a clear commitment from NavTalk for production and high-end compute. Use of an NVIDIA driver that supports RTX 5090 (e.g. 5xx series) and a common Linux distro (e.g. Ubuntu 22.04 LTS or newer) is recommended.

2. Technology Stack and Adaptation

NavTalk’s runtime on 5090 Linux is separately selected and validated, and is maintained separately from environments for older GPUs (e.g. CUDA 11.8), to avoid wrong installs or mixing and to simplify environment isolation and reproducibility.

2.1 Core Runtime (5090-specific)

The table below lists officially verified software versions for NavTalk on 5090, for ops and integration reference. Python is the runtime; CUDA is NVIDIA’s GPU compute platform; PyTorch is the main framework for AI models; mmcv / mmdet / mmpose are the vision libraries used for face and pose (e.g. for digital humans).

NavTalk maintains a dedicated dependency list for 5090, listing the above libraries and versions and their relationship to TensorFlow, CUDA 12.8, NumPy, etc., separate from older GPU environments, reflecting dedicated 5090 support and maintainability.

2.2 Compatibility for 5090 Architecture

RTX 5090 uses the new Blackwell architecture (compute capability 9.0). Some vision libraries may not ship ready-made binaries for 5090. Compatibility has been verified and adapted for the 5090 architecture so that face, pose, and related capabilities run correctly on 5090 and full usability is ensured.

2.3 Inference and Model Management

▪️ NavTalk’s lip-sync core is based on MuseTalk 1.5 (a widely used high-quality lip-sync model) and runs on 5090 with the PyTorch 2.7 + CUDA 12.8 stack above.

▪️ NavTalk provides a unified GPU and model management layer: models are loaded on demand, and multi-task contention for the GPU is avoided, improving stability for multi-service or multi-GPU setups and long-running use on 5090.

All versions and adaptations above are validated and reflect reproducible, deliverable engineering support, not just “theoretical” compatibility.

3. Product Value and Use Cases

▪️ Real-time performance and quality: On 5090, NavTalk can fully use the new compute to achieve real-time lip-sync at 30+ fps and support higher resolution and multi-session concurrency, suitable for digital humans, virtual hosts, and live interaction where latency and quality matter.

▪️ Service shapes: On 5090 Linux, NavTalk offers real-time lip-sync API, video lip-sync API, and digital human avatar API, among others, for live, recorded, and interactive use; the real-time lip-sync API is optimized for low latency and streaming.

▪️ Production-ready: Concurrency, quality enhancements (e.g. face enhancement, mouth sharpening), GPU options, and output directories are configurable, so NavTalk can integrate with your existing business systems, storage, and monitoring.

Supporting NavTalk on 5090 Linux is therefore full production-ready support for the latest compute, not merely “it runs,” and supports evaluation and rollout decisions.

4. 5090 Concurrency and Response

▪️ Concurrent sessions: RTX 5090 has about 32 GB VRAM; from a VRAM perspective it can support about 2 concurrent sessions (actual limits also depend on system RAM and other factors; load testing on the target configuration is recommended).

▪️ Response: 5090 has more compute than 4090/3090, so per-frame inference is faster; with a single session, GPU utilization at idle is 0%, which supports low latency and real-time lip-sync.

5. Summary

▪️ NavTalk is officially supported and runs fully on NVIDIA RTX 5090 + Linux. For 5090, NavTalk defines runtime versions (e.g. Python 3.10, CUDA 12.8, PyTorch 2.7), a 5090-specific dependency list, and recommended versions for face/pose libraries, and completes end-to-end adaptation and validation from drivers and frameworks to the inference engine.

▪️ Compatibility for the 5090 (Blackwell) architecture is in place; when prebuilt packages are not available, building from source on 5090 is also supported.

▪️ Concurrency and response: On 5090, about 2 concurrent sessions are supported from a VRAM perspective; higher compute improves low-latency and real-time lip-sync.

This document is a product statement of NavTalk’s 5090 Linux support, for external communication and technical evaluation.

This document describes NavTalk’s official support for RTX 5090 on Linux; see current product documentation for up-to-date details.