General Director Jensen Juan went on stage on CES in Las Vegas to demonstrate new offers of equipment and software, which cover everything from personal AI supercomputers to next-generation gaming cards.
The largest announcement of NVIDIA: Project Digits, AI personal supercomputer in the amount of $ 3,000, which is packaged by Petaflop computing power in the box the size of a desktop.
Built around the new – and still secret – Grate 10 Blackwell Superchip, this machine can process AI models with parameters up to 200 billion when attracting power from a standard outlet.
For heavier workloads, users can connect two blocks to combat models up to 405 billion parameters.
For context, the largest LLAMA 3.2 model, the most advanced LLM with an open source from META, has 405 billion parameters and cannot be launched on consumer equipment.
Until now, it has been required by about 8 NVIDIA A100/H100 superchites, each of which cost about 30 thousand dollars, which is more than 240 thousand dollars. The United States is only for hardware processing.
Two of the new Ai AI Nvidia supercomputers will cost 6 thousand dollars and will be able to perform the same quantum model.
“AI will be mainstream in each application for each industry. Using Project Digits Superchip Grace Blackwell comes to millions of developers, ”Jensen Juang, general director of NVIDIA, said in an official message on the blog. “The placement of the AI supercomputer on the tables of every scientists, the researcher of AI and the student gives them the opportunity to involve and form the age of AI.”
For those who love technical details, the GB10 chip is a significant engineering achievement that gave birth in cooperation with MediaTek.
The Ship system combines the latest NVIDIA GPU architecture with 20 powerful ARM nuclei connected via NVLINK-C2C Interconnect.
Each block of numbers has 128 GB of single memory and up to 4 TB of NVMe storage. Again, for the context, the most powerful graphic processors today are about 24 GB of VRAM (memory necessary for launching AI models), and the H100 Superhip starts at 80 GB of VRAM.
NVIDIA plans for the dominance of AI agents
Companies seek to deploy AI agents, and Nvidia knows this, therefore, therefore, he has developed Nemotron, a new family of models that are delivered in three sizes, and announced their expansion today with two new models: Nvidia NIIM for summing up video and understanding and understanding and understanding and understanding and understanding and understanding and understanding and understanding and understanding And understanding and understanding and understanding and understanding and understanding and understanding and understanding and understanding and understanding and understanding and understanding and understanding and understanding and understanding. NVIDIA COSMOS for providing Nemotron Vision capabilities – the ability to understand visual instructions.
Until now, LLM has been only on the basis of the text. Nevertheless, the models succeeded in the following instructions: chat, challenges of functions, coding and mathematical tasks.
They are available both with both the hugging person and on the NVIDIA website, with access to the enterprise through the company’s software platform of the AI Enterprise company.
Again, for the context, in the LLM arena, Llama Nemotron 70b from NVIDIA occupies a higher rating than the original LLAMA 405B developed by META. It also exceeds different versions of Claude, Gemini Advanced, Grok-2 Mini and GPT-4O.
The NVIDIA agent is also currently associated with infrastructure. The company announced partnerships with large agency suppliers of technical technologies such as Langchain, Lmamaindex and Crewai to build drawings on Nvidia Ai Enterprise.
These finished templates deploy specific tasks that facilitate the developers to create highly specialized agents.
The new PDF-to-PODCAST plan seeks to compete with Google Notebooklm, while another plan helps to create a video of the search and consolidated agents. Developers can check these drawings through the new NVIDIA LaUNCHAILS platform, which provides prototyping and deployment of one click.
Gamers, rejoice! New GEFORCE RTX 5000 cards is a beast
Nvidia has retained its game ads for recently, publishing a very expected GeForce RTX 5000 series. The flagship RTX 5090 contains 92 billion transistors and provides 3352 trillion operations per second, which gives the performance of current RTX 4090. The entire composition has the fifth -generation tensor kernels and RT -Fourth generation.
New cards are introduced by DLSS 4, which can increase the frame rate to 8x, using AI to generate several frames for rendering. Blackwell, the AI engine, arrived for PC-Gamers, developers and creative people, ”said Jensen Juang,“ the unification of neural rendering and trace of the beam, based on artificial control, is the most significant computer graphic innovation, as we presented a programmable shading of 25 years back. “
New cards also use the models of transformers for super-resolution, promising a highly realistic graphics and much greater performance at their price, not cheap, by the way: $ 549 for RTX 5070, from 5070 Ti $ 5080 to 999 US dollars and 5090 according price of 1999 dollars.
If you do not have that kind of money and you want to play, do not worry.
AMD also announced today about its Radeon RX 9070 series. Cards are built on the new RDNA 4 architecture using 4 -NM production process and special specialized AI accelerators to compete with NVIDIA tensor nuclei.
While complete specifications remain under the extensive, the last Ryzen AI chips from AMD already reach 50 peaks with peak performance.
Unfortunately, NVIDIA is still the king of AI applications thanks to its CUDA technology, patented architecture of AI NVIDIA.
To cope with this, AMD provided partnerships with HP and ASUS to integrate systems, and more than 100 brands of corporate platforms will use AMD Pro Technology until 2025.
It is expected that Radeon cards will appear on the market in the first quarter of 2025, giving Nvidia an interesting battle both in games and in acceleration of artificial intelligence.
Edited by Sebastian Sinclair
