Nvidia introduces its AI systems into data centers and what it calls AI factories around the world and companies announced Today, his Blackwell chips lead to AI comparative tests.
Nvidia and its partners accelerate the training and implementation of the up-to-date generation AI applications that exploit the latest progress in training and inference.
The NVIDA Blackwell architecture is built to meet increased performance requirements for these up-to-date applications. In the latest MLPERF-12 training round. Since the introduction to the comparative test in 2018, the AI NVIDIA platform has ensured the highest performance on a scale at every comparative point and supplied each result in the most hard model of a huge language (LLM): LLME 3.1 405B.
The NVIDIA platform was the only one that submitted the results in each height of MLEPERF Training V5.0 – emphasizing its exceptional performance and versatility in a wide range of loads AI, LLM, recommendation systems, multimodal LLM, detection of facilities and a chart of neural networks.
Applications on a scale on a scale were used two AI supercomputers powered by the NVIDIA Blackwell: Type, built using the NVIDIA GB200 NVL72 and NYX standing indicator systems, based on NVIDIA DGX B200. In addition, NVIDIA cooperated with Coreweave and IBM to send GB200 NVL72 results using a total of 2496 GPU Blackwell and 1248 NVIDIA Grace processors.
In the up-to-date level of the Llam 3.1 405B test test, Blackwell provided 2.2 times higher performance compared to the architecture of the previous generation on the same scale.
At Lama 2 70B Lora, refining the NVIDIA DGX B200 Systems reference points, powered by eight GPU Blackwell, provided 2.5 times more performance compared to the application using the same number of GPU in the previous round.
These performance steps emphasize the progress in Blackwell architecture, including high-density liquid racks, 13.4 TB of consistent memory for a stand, Nvidia Nvlink and Nvidia Nvlink Switch Switch Interconnect Technologies for scaling and nvidia quantum-2 infiniband networking. In addition, innovations in the Pile of NVIDIA NEMO Framework software raising the bar for the up-to-date generation multimodal LLM training, crucial for introducing AI agency applications on the market.
One day these agency agency applications will operate in AI AI AI-Sailors. These up-to-date applications will create tokens and valuable intelligence that can be used to almost every industry and academic domain.
The NVIDIA Data Center platform includes GPU, CPU, brisk fabrics and network creation, as well as a wide range of software, such as the NVIDIA Miracle-X, NEMO Framework, NVIDIA TENSIRT-LLM and NVIDIA Dynamo libraries. This highly tuned team of hardware and software technology enables organizations to train and implement models faster, dramatically accelerating time to value.
The NVIDIA partner ecosystem intensively participated in this MLPERF round. In addition to submitting with Coreweave and IBM, other convincing entries came from ASUS, Cisco, Giga Computing, Lambda, Lenovo Quanta Cloud Technology and Supermicro.
The first MLEPERF training applications using the GB200 were developed by the MLCcommons Association with over 125 members and associated entities. Over time, the training record ensures that the training process produces a model that meets the required accuracy. And its standardized comparative rules ensure comparisons of apple performance with apples. The results are reviewed before the publication.
Basics of training tests
Dave Salvator is someone I knew when he was part of the technological press. Now he is the director of accelerated computer products in an accelerated computer group in NVIDIA. In the press briefing Salvator noticed that the general director of NVIDIA, Jensen Huang, talks about this concept of scaling regulations for artificial intelligence. They include initial training in which you basically teach AI model knowledge. It starts from scratch. Salvator said it is a weighty computing elevator, which is the spine of artificial intelligence.
From there, Nvidia goes to scaling after training. At this point, the models go to school, and it is a place where you can do things such as tuning, in which you introduce a different set of data to teach a previously trained model, which has been trained to some extent to give him additional knowledge of the domain about a specific set of data.
And finally, there is scaling or transient reasoning, and sometimes called long thinking. Agentic AI is the second date that passes. It is artificial intelligence that can think, reason and solving problems in which you basically ask a question and get a relatively basic answer. Testing time and reasoning can actually act on much more complicated tasks and provide a luxurious analysis.
And then there is also generative artificial intelligence that can generate content if necessary, which may contain translations summarizing the text, but also visual content and even audio content. There are many types of scaling that continues in the AI world. In the case of comparative tests, NVIDIA focused on pre -training results and after training.
“This is where AI begins what we call the investment phase of artificial intelligence. And then, when you start to apply and implement these models, and then generating these tokens, you just start to get a return on investment in artificial intelligence,” he said.
Benchmark Mlperf is in the 12th round and comes from 2018. A consortium, which has over 125 members and was used for both application and training tests. The industry perceives reference as solid.
“I am sure that many of you are aware, sometimes claims for performance in the world of artificial intelligence can be a little wild west. Mlperf is trying to bring some orders of this chaos,” said Salvator. “Everyone must do the same work. Everyone is kept on the same standard in terms of convergence. And after sending the results, these results are then checked and checked by all other applications, and people can ask questions and even question the results.”
The most intuitive record around training is how long the training of the AI model has been trained to the so -called convergence. This means achieving a specific level of accuracy. Salvator said it is a comparison of apples with apples, and takes into account constantly changing loads.
This year, there is a up-to-date load on Llam 3.140 5b, which replaces the chatgpt 170 5b load, which was previously at the reference point. In comparative tests, Salvator noticed that Nvidia had many entries. AI Nvidia GB200 NVL72 factories are fresh from production factories. From one generation of systems (funnel) to the next (Blackwell), Nvidia has recorded a 2.5 -fold improvement for image generation results.
“We are still quite early in the Blackwell product life cycle, so we fully expect more performance from Blackwell architecture over time, because we are still improving our software optimization and as new, to be honest, heavier loads work on the market,” said Salvator.
He noticed that Nvidia was the only company that made entries for all reference points.
“The great performance we achieve comes from a combination of things. It is our fifth generation Nvlink and NVSWitch that provide up to 2.66 times more performance, along with others only the general architectural goodness in Blackwell, along with our ongoing software optimizations that allow this performance,” said Salvator.
He added: “Due to the heritage of Nvidia, we have been known for the longest time, because these GPU guys. We certainly create great GPU, but we have gone from being a chip company to be a system company with such as our DGX servers, now building whole stands and data centers with our rack designs, which are now reference projects, which are now reference projects. Infrastructure. which we currently call AI factories.