news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Earlier, there were reports that the delivery of the new generation of GPUs based on NVIDIA Blackwell architecture was delayed, but the official neither admitted nor denied it. Instead, it quietly took out a server system based on the new generation GB200 GPU+GPU super chip and demonstrated it.

GB200 Grace Blackwell includes two B200 GPUs and one Grace CPU, with a total power consumption of up to 2700W.

Among them, B200 has 204 billion transistors, 8096-bit 192GB HBM3E memory (bandwidth 8TB/s), performance reaches 9PFlops (9 quadrillion times per second), and power consumption reaches 1000W for the first time.

The server displayed by NVIDIA is like a work of art, with blue and copper cables, a large number of ultra-thick diameter copper tubes, and super-large heat sinks forming a huge monster.

NVIDIA did not disclose the specific configuration, only saying that there are two versions with and without NVSwitch, with a total bandwidth of up to 14.4TB/s.

Only one data point was given for performance: the AI ​​reasoning performance of the Llama 3.1 70 billion parameter large model is 1.5 times higher than that of the H200.

How to dissipate heat? NVIDIA revealed that it is developing a hot water cooling system that can reduce power consumption by 28%.

However, NVIDIA still has not given a shipping date for Blackwell GPU servers, but it is estimated to be in the fourth quarter of this year.

At the upcoming Hot Chip 2024 conference, NVIDIA will explain the Blackwell GPU in detail and further disclose news about future Vera CPUs and Rubin GPUs.