Pular para o conteúdo
Menu
Mix Vale

Nvidia’s free update speeds up LLMs by up to 40% on RTX cards with new technology

Luís Henrique Costa
Por 7 min de leitura
Compartilhe
Nvidia
Foto: Nvidia - Foto: Hepha1st0s / Shutterstock.com

Nvidia has announced a free software update that promises to significantly increase artificial intelligence performance on computers equipped with its RTX graphics cards. The suite of optimizations directly benefits users running large language models (LLMs) locally, as well as accelerating generative content creation tasks. The improvements, which include support for new precision formats to reduce VRAM memory consumption, solidify the position of RTX GPUs as the leading platform for AI workloads in the consumer environment, delivering greater speed and efficiency.

Publicidade

The new package combines advances in processing speed and graphic resource management, continuing a series of improvements that the company has been implementing since 2023. With the update, PC users with RTX cards gain immediate access to these tools at no additional cost, democratizing the use of advanced AI models that previously required specialized hardware or access to cloud services. The initiative reinforces Nvidia’s strategy of differentiating its GPUs from integrated NPUs (Processamento Neural Units), which are generally limited to more basic tasks.

The update breaks down into several key components that impact different aspects of generative artificial intelligence, from running chatbots to creating high-resolution videos. The optimizations are designed to work seamlessly with the Windows operating system and popular tools in the AI ​​ecosystem, ensuring fluid implementation for developers and content creators who rely on local processing power for their workflows.

Publicidade
Nvidia
Nvidia – Jack Hong/shutterstock.com

Expressive speedup in language models

One of the pillars of this update is the increase in speed for executing large language models. Testes internals performed by Nvidia indicate performance gains of up to 40% on popular LLMs such as Nemotron Nano V2 and several open source variants of the GPT family. Esse performance jump is a direct result of optimizations implemented in the TensorRT-LLM library, which now allows faster and more efficient inference.

In practice, users will notice shorter response times in text assistants and chatbots running locally. Essa improvement especially benefits developers and creators who integrate LLMs into their daily routines, enabling more agile iterations and a more fluid user experience. The acceleration applies to both RTX 40 series cards and RTX 30 series models, expanding the scope of the benefit to a broad user base.

[[MVG_PROTECTED_BLOCK_0]

These advances represent the continuation of an effort started by Nvidia in 2023, when the first accelerations via TensorRT-LLM were introduced. The company continues to refine its software to extract the maximum potential from the Tensor Cores present in its GPUs, ensuring that even complex AI tasks can be executed efficiently on consumer hardware.

New NVFP4 technology optimizes VRAM usage

Another highly relevant technical novelty is the introduction of native support for the NVFP4 precision format. Essa quantization technology allows AI models to be compressed by up to 60% compared to traditional versions using the BF16 format. Compression drastically reduces the space the model occupies in video memory (VRAM), one of the most critical resources for running large LLMs.

This compression works by transferring part of the processing load to the system’s RAM memory, freeing VRAM for other simultaneous activities. In image creation tools like ComfyUI, for example, using NVFP4 can result in gains of up to 4.6 times in performance for pipelines that use models like Flux.1 and Flux.2. Isso means users can generate high-quality images faster and more efficiently.

Publicidade

The reduction in VRAM consumption also democratizes access to larger and more complex models. Usuários with video cards with smaller memory capacity, including models from previous generations, gain the possibility of running models that would previously have been unfeasible. Essa optimization extends the life of existing hardware and pushes the boundaries of AI experimentation on mainstream desktops.

Support also extends to the NVFP8 format, which offers a different balance between compression and precision, making it particularly useful in other applications such as video generation. The flexibility offered by these new formats allows developers and users to choose the best configuration for each type of task, optimizing the use of system resources according to need.

Advances in video generation with the LTX-2 model

In the field of audiovisual content creation, Nvidia has deepened its collaboration with Lightricks to optimize LTX-2, one of the leading open source models for video and audio generation. Este model is capable of producing clips at native resolution up to 4K at 50 frames per second, with fully synchronized audio, a computationally intensive task that now becomes more accessible.

Thanks to support for the NVFP8 precision format, the LTX-2 model can achieve double the generation speed on the latest RTX cards. On compatible hardware, a high-quality video can be generated in approximately 20 seconds, a remarkably short time for this task. The technology stands out for its ability to create long-form content with integrated audio, offering creators a powerful tool for fast, local production while maintaining full creative control and privacy over data.

RTX Video Super Resolution for generative content

The popular RTX Video Super Resolution (VSR) feature, which uses AI to improve the quality of videos in real time, has been expanded to also cover videos created by generative artificial intelligence. The tool upscales content from lower resolutions, such as 720p, to 4K, applying significant improvements to the level of detail and sharpness of the image.

The integration is expected to arrive on the ComfyUI platform in February 2026. With this optimization, the entire process of generating and upscaling a 10-second 4K clip, which could take up to 15 minutes, will be reduced to just 3 minutes. Essa acceleration directly benefits content producers who need high-resolution output quickly and efficiently, leveraging the specific hardware capabilities of RTX GPUs to maintain superior visual quality.

RTX ecosystem and the democratization of local AI

This update consolidates Nvidia’s vision to position the RTX platform as the definitive and most complete solution for developing and running artificial intelligence on-premises. By delivering performance gains of up to 40% in LLMs, reducing VRAM consumption by up to 60% with new precision formats like NVFP4, and dramatically accelerating 4K video generation, the company not only improves the user experience, but also establishes a clear differentiation from competing hardware solutions like integrated NPUs, which are typically limited to simpler, less computationally demanding AI tasks. The Nvidia strategy is based on a robust ecosystem that combines powerful hardware, such as the Tensor Cores, and constantly evolving software, such as the TensorRT-LLM libraries. Isso empowers a wide range of users, from developers and researchers to content creators and enthusiasts, to explore the frontiers of generative AI right on their desktops. Running on-premises offers crucial advantages such as reduced latency, full control over data and privacy, and the elimination of dependence on cloud services and their associated costs. With these tools, Nvidia is effectively democratizing access to cutting-edge AI technologies, enabling innovations to emerge from anywhere.

Improved integration with the ComfyUI platform

The ComfyUI platform, one of the most popular node-based environments for building generative AI workflows, receives Nvidia-specific optimizations in this update. The tool directly benefits from NVFP4 format support in image models like Flux, allowing users to configure complex pipelines with a much lower demand on graphics resources.

Ongoing collaboration between Nvidia and developers in the open source ecosystem ensures that software improvements are quickly integrated and made available to the community. Essas changes simplify experimentation with large models and advanced workflows, making the technology more accessible to a broader audience.

Practical benefits for content creators

For professionals working with image and video creation, speed gains translate into faster iterations and a more dynamic workflow. The ability to generate professional clips with the LTX-2 model without relying on cloud services and the reduction in VRAM consumption allows multitasking with multiple models active simultaneously, increasing productivity and creative flexibility.

Publicidade

Luís Henrique Costa é jornalista formado pela FMU, editor-chefe e fundador do Mix Vale, portal de notícias em atividade desde 2011. Responde pela direção editorial do veículo, que cobre Brasil e mundo em economia, benefícios sociais, esportes, tecnologia, ciência e entretenimento. Atua na área há mais de 15 anos e lidera uma equipe de cerca de 15 colaboradores.

Notícias relacionadas

Mais lidas

  1. Ciência

    Meteoro no Nordeste: bólido cruza o céu de cinco estados às 21h34 e mobiliza buscas por fragmentos na Paraíba

  2. Caixa Econômica Federal

    Pé-de-Meia paga parcela de R$ 1.000,00 a aprovados no ensino médio de 2026 entre fevereiro e março de 2027

  3. Política

    Com obstrução parcial no intestino, Jair Bolsonaro é internado no hospital DF Star em Brasília

  4. INSS

    Recuperação da senha Gov.br pelo celular exige reconhecimento facial ou código por SMS para o Meu INSS

  5. EUA

    Terremoto em Los Angeles atinge magnitude 4,2 e causa tremores leves em Commerce na madrugada

  6. Matéria Publicitária

    MineMV é o novo jogo grátis de construir com blocos; veja como jogar e liberado para todas as idades

  7. Política

    Regulamentação da escala 5×2 pelo governo encarece cálculo da hora extra a partir de 16 de dezembro

  8. Santa Catarina

    Acidente de trânsito na SC-108 envolve carro e moto e deixa motociclista morto em São Ludgero

  9. Mundo

    Jornalista italiano Corrado Augias morre aos 91 anos; presidente da Itália lamenta a perda

  10. Loterias

    Rendimento de R$ 40 milhões da Mega-Sena 3088 chega a R$ 337 mil por mês no Tesouro Direto