The entrance of NVIDIA in the computing market through AI At the server level, it has allowed the company to reach historical highs, both in sales and in developed technologies, since they have managed to become the company with the highest value in the world thanks to the development of components related to artificial intelligence. But obviously there are always ways to improve the power that the hardware is capable of achieving, and this time they have created a superchip which combines multiple B200 GPUs along with a pair of Grace CPUs to deliver the highest possible data center performance.
Data centers have high-performance systems that allow a large amount of information to be processed simultaneously, something that is increasingly increasing due to the development of new technologies that require computing power much higher than what we could see ago. only four years. For this reason, those companies that have specialized hardware for a technology that allows data processing to be faster and easier are the ones that currently lead the market, being the main representative of this case. NVIDIA thanks to your chips created for AI.
NVIDIA’s new AI processor has four B200 GPUs and two Grace CPUs fused together
There are more and more advances that we see in the computing sector that allow us to manage in a much more efficient way the large amount of data that is generated daily on servers, and in this case the arrival of artificial intelligence has managed to improve this aspect a lot. Now NVIDIA has presented a new chip that will allow for even higher performance, as it combines several of its most important technologies to surpass the power offered by the GB200 Grace Blackwell, which already incorporated two high-performance GPUs, but the new one includes the double.
And it is that he GB200 Grace Blackwell NVL4 Superchip It is a variant that incorporates a total of four Blackwell B200 GPUs connected together via NVLink, along with two ARM-based Grace CPUs, all combined on one motherboard (which is why we don’t know why the company calls it a “superchip” ). The new GB200 NVL4 It manages to offer 2.2 times higher performance in simulation, 1.8 times higher in training and 1.8 times higher in inference compared to its predecessor, the GH200 NVL4 Grace Hopper Superchip.
In addition to this, the company has also introduced a new dual-slot graphics card, the NVL H200intended for data centers that has air cooling and PCIe 5.0 connectivity, which should allow reaching 128 GB/s. This model is specifically designed to present a continuous airflow that circulates from right to left, without blower fans, in addition to being optimized for rack mounting solutions. Its performance will be slightly lower than that offered by the H200 GPUbut the company assures that the NVL H200 is far superior to the one it intends to replace, the NVL H100 since it has 1.5 times more memory capacity and 1.2 times more memory bandwidth.
