NVIDIA’s GeForce RTX 5090 Outperforms in Inference on DeepSeek’s R1 AI Models, Far Exceeding AMD’s RX 7900 XTX

NVIDIA’s latest graphics card, the GeForce RTX 5090, is turning heads with its remarkable performance on the DeepSeek R1 models, far outpacing AMD’s RX 7900 XTX. This can largely be attributed to NVIDIA’s cutting-edge fifth-generation Tensor Cores, which really make a difference in the processing speed.

### Effortless Access to DeepSeek’s Reasoning Models with the New RTX GPUs, Delivered with Exceptional Performance

It looks like the world of consumer GPUs is becoming a top contender for running sophisticated language learning models (LLMs) on personal computers. NVIDIA and AMD are both pushing the boundaries to create optimal conditions for these powerful tasks. Not too long ago, AMD demonstrated the capabilities of their RDNA 3 flagship GPU using the DeepSeek R1 LLM model, capturing a lot of attention. In a bold counter-move, NVIDIA has now showcased the inference benchmarks on their latest RTX Blackwell GPUs, revealing that the GeForce RTX 5090 holds a significant advantage in performance.

When tested across various DeepSeek R1 models, the GeForce RTX 5090 displayed a noticeable lead over AMD’s Radeon RX 7900 XTX and even NVIDIA’s previous generation GPUs. Particularly, it achieved throughput of up to 200 tokens per second with Distill Qwen 7b and Distill Llama 8b models, doubling the performance of AMD’s RX 7900 XTX. This indicates a strong future for AI tasks on NVIDIA GPUs, especially with the robust “RTX on AI” support gearing up, paving the way for more advanced consumer AI applications.

NVIDIA is making it easier than ever for enthusiasts wanting to explore DeepSeek R1 on their RTX GPUs. They’ve even put together a handy guide in a blog post—it’s as straightforward as using any online chatbot. Here’s a quick rundown of how you can dive in:

> For developers eager to explore and create custom AI agents, NVIDIA is offering their massive 671-billion-parameter DeepSeek-R1 model as a NIM microservice preview at build.nvidia.com. This microservice can churn out up to 3,872 tokens per second on a single NVIDIA HGX H200 system.
>
> You can experiment with the API, which is soon expected to be available as a downloadable NIM microservice, integrated within the NVIDIA AI Enterprise software ecosystem.
>
> The DeepSeek-R1 NIM microservice streamlines deployment across industry-standard APIs. Companies that prioritize security and data privacy will appreciate running this on their preferred accelerated computing platforms.
>
> – NVIDIA

With NVIDIA’s NIM platform, developers and tech enthusiasts can seamlessly test the AI model on their own hardware. This not only ensures that your data remains secure, but also optimizes performance, assuming your setup meets the necessary hardware requirements.