Technological innovation currently comes from AI, and with the new architecture Blackwell, NVIDIA seeks to lead the new era of computing, two years after the launch of NVIDIA Hopper, becoming the most powerful chips in the world. The presentation took place during the GTC, and has managed to quickly capture the attention of all people because it is an architecture that seeks to promote both accelerated computing and AI generative.
From NVIDIA They assure that this new architecture will be able to meet the expectations of the most demanding companies, since it will allow a large number of organizations to create and execute AI real-time generative on language models with billions of parameters. On the other hand, the cost will be reduced considerably, since the company that created this chip assures that its new architecture will have a cost and energy consumption 25 times lower.
NVIDIA, leader in the AI ​​sector
Over the last two years, since the Hopper architecture was released, we have seen how NVIDIA It has managed to be among the positions of the largest organizations worldwide, with extremely high income. This in itself is the reason for having created what was until now the most powerful chip in the world at the level of a commercial solution for products related to AImaking the majority of its competitors unable to even come close to the power that the Hoppers are capable of offering.
Now, two years later, they are back at it to present the chips Blackwell which in addition to having improved in terms of structure and power, also incorporate six new computational acceleration technologies that will allow the company to make its way into many more fields. Some of the most famous companies in the world have already decided that they will incorporate these chips, since they see it as an opportunity to develop new technologies and products in a much more advanced way.
Together, the new six technologies included in this new architecture presented by NVIDIA allow the training of AI and the use of real-time LLM for models with up to 10 billion parameters.
What Blackwell architecture introduces
As we have indicated before, one of the key factors of this superchip is that it incorporates six new technologies that are key to understanding its use, since it is directly aimed at offering higher data processing speeds, which is perfect for both accelerated computing and generative AI.
This new chip includes 208,000 million transistors, being a GPU manufactured through a customized process called TSMC 4NP, which also integrates the frameworks NVIDIA TensorRT-LLM and NeMo Megatron, which together with the fifth generation NVLink, will guarantee the highest possible speed, thanks to a bidirectional performance of 1.8 TB/s.
On the other hand, chips with architecture Blackwell include an engine dedicated to reliability, availability and maintainability, as well as a series of advanced security features that help protect the models of AI and customer data without reducing performance, making it reliable, both inside and out.
And finally, it also includes a dedicated decompression engine, which by supporting all current formats, allows you to accelerate database queries, thus offering the highest possible performance in issues related to data analysis.
The multi node system NVIDIA GB200 NVL72, will be one of those that ensures providing maximum power with the lowest possible consumption, since it combines 36 Grace Blackwell Superchips, including 72 GPUs Blackwell and 36 Grace CPUs interconnected via fifth-generation NVLink. Provides up to 30 times more performance than the same number of GPUs NVIDIA H100 Tensor Core for LLM inference workloads and reduces cost and power consumption by up to 25x.
