NVIDIA says its Blackwele chips lead standards in AI LLMS training

Nafidia AI chips are offered to data centers and what you call artificial intelligence factories around the world, and the company Declare Today, Blackwell chips lead to artificial intelligence standards.

NVIDIA and its partners speed up training and publish AI applications from the next generation, which use the latest developments in training and reasoning.

Nvida Blackwell architecture is designed to meet the increasing performance requirements of these new applications. In the last round of MLPERF-twelfth place since the presenter of the standard in 2018-the NVIDIA AI platform presented the highest wide performance on each standard and operated each result presented on the most difficult major linguistic model (LLM) -Llama 3.1 405b.

The NVIDIA platform was the only platform that presented the results on each MLPERF V5.0 training standard – which confirms its exceptional performance and its diversity through a wide range of artificial intelligence work burden, stretch LLMS, recommendation systems, multimedia LLMS, and discovering neurons and networks graphs.

The presentations widely used two super -computers from the NVIDIA Blackwell: Tyche, which were designed using NVIDIA GB200 NVL72, and NYX, based on NVIDIA DGX B200 systems. In addition, NVIDIA cooperated with Coreweave and IBM to provide GB200 NVL72 results using a total of 2,496 Blackwell graphics processing units and 1,248 Nvidia Grace CPU.

On the new Lama 3.1 405B standard, Blackwell has performed a 2.2 -time performance compared to architecture of the previous generation on the same range.

On Llama 2 70b Lora the correct standard, NVIDIA DGX B200, which works with eight Blackwell graphics processing units, performed more than 2.5 times more compared to using the same number of graphics processing units in the previous round.

Performance leaps highlight these developments in the Blackwell structure, including high-density liquid shelves, 13.4 terabytes of coherent memory for each shelf, Nvidia Nvlink, and NVIDIA Nvlink Switching to expand the NVIDIA-2 Infiniband networks. In addition, innovations are raised in a staple of NVIDIA Nemo Framework for the next generation of LLM, which is very important to provide AIC applications to the market.

These AI applications will one day run in artificial intelligence factories-AI Agenic Economics Engines. These new applications will produce the distinctive codes and valuable intelligence that can be applied to almost every field of industry and academy.

It includes the Nvidia Data Center GPU and CPU, high-speed fabrics and networks, as well as a wide range of programs such as Nvidia Cuda-X Libraries, Nemo Framework, Nvidia Tensorrt -lm and Nvidia Dynamo. This very seized group enables the technologies of devices and programs to train and publish models more quickly, and to speed up time significantly.

The NVIDIA’s ecosystem is widely shared in this Mlperf tour. In addition to applying with Coreweave and IBM, other convincing offerings were from ASUS, Cisco, Giga Computing, Lambda, Lenovo Quanta Cloud Technology and SuperMicro.

The first MLPERF training introductions were developed using GB200 by MLCOCONMONS Association with more than 125 members and subsidiaries. The training time scale ensures that the training process produces a model that meets the required accuracy. Standard standard operating rules include performance comparisons of apples. The results reviewed before the publication.

The basics in training standards

Dave Salvator is a person I knew when he was part of the Technology Press. He is now the director of accelerating computing products in the accelerated computing group in NVIDIA. At a press conference, Salvator indicated that the CEO of Nvidia Jensen Huang talks about this idea of ​​AI’s scaling laws. It includes training before training, as it mainly teaching Amnesty International’s international knowledge. This starts from scratch. Salvator said it was a heavy arithmetic elevator, the backbone of Amnesty International.

From there, NVIDIA moves to scaling after training. This is where the models go to school, and this is where you can do things like accurate synthesis, for example, as it brings a different data collection to teach a pre -trained model that has been somewhat trained, to give him an additional field of your data set.

Then finally, there is a scaling time or thinking test, or is sometimes called long thinking. The other term passes through it is the agent of artificial intelligence. It is Amnesty International, which can actually think, mind and problem, as it raises a basic question and obtains a relatively simple answer. Test at the time of scaling and thinking in reality can work on more complex tasks and provide rich analyzes.

Then there is also an Amnesty International Obstetrician that can create content on the basis as needed that can include text summary translations, but also visible content and even audio content. There are many types of scaling that occur in the world of artificial intelligence. For standards, NVIDIA focused on pre -training and post -training results.

He said: “This is the place where artificial intelligence begins with what we call the stage of investment in artificial intelligence. Then when you enter into reasoning and publish these models and then generate these symbols, as you start getting your return on your investment in artificial intelligence.”

MLPERF standards are located in the twelfth round and dates back to 2018. The consortium, which supports it, contains more than 125 members and has been used in inference and training tests. The industry sees strong standards.

“I am also sure that many of you are familiar with, sometimes performance claims in the world of artificial intelligence can be a little brutal West. Mlperf seeks to provide some orders to that chaos.” “Everyone must do the same amount of work. Each person is kept in the same standard in terms of rapprochement. Once the results are presented, these results are reviewed and then examined by all other applicants, and people can ask questions and even the challenge results.”

The most easy scale on training is the time that the trained artificial intelligence model training takes on the so -called rapprochement. This means hitting a specific level of accuracy. Salvator said it was compared to apples to apples, and it takes into account the constant change of work burdens.

This year, there is a new Llama 3.140 5B work burden, which replaces the Chatgpt 170 5B work burden that was previously in the standard. In the standards, Salvator noted that NVIDIA had a number of records. New NVIDIA GB200 Nvl72 AI factories from manufacturing factories. From one generation of chips (Hopper) to the following (Blackweell), NVIDIA has seen a 2.5 -time improvement for the results of the image generation.

“We are still somewhat early in the Blackwell product life cycle, so we expect completely to get more performance over time from the Blackweell brown, as we continue to improve our program improvements, and with the new and heavier work burdens to the market,” Salvator said.

He pointed out that NVIDIA was the only company that provided all criteria.

“The great performance we achieve comes through a mixture of things. It is the Fifth Generation and Nvswitch to provide up to 2.66 times, along with the other general architectural good in Blackweell, along with the continuous software improvements that make this performance possible,” said Salvator.

He added: “Because of the heritage of Nvidia, we were known for the longest time of such players in the graphics processing unit. Infrastructure, which we now refer to as artificial intelligence factories.