"Informed AI News" is an publications aggregation platform, ensuring you only gain the most valuable information, to eliminate information asymmetry and break through the limits of information cocoons. Find out more >>
NVIDIA's Blackwell Platform: A Milestone in AI Computing
- summary
- score
NVIDIA's Blackwell platform, a powerhouse for AI, comes with a hefty price tag. The GB200 NVL72 server rack system, a beast with 72 GPUs, costs $3 million. It's part of a $100 billion investment, involving 25,000 people.
The demand for these high-end AI servers remains strong, despite the cost. NVIDIA plans to ship 60,000 to 70,000 of these racks by 2025, aiming for $2100 billion in revenue. Each rack, packed with 2080 billion transistors, offers 20 petaflops of computing power.
Training complex AI models, like a 1.8 trillion parameter model, requires massive resources. 8000 Hopper GPUs or 2000 Blackwell GPUs are needed, with significant differences in power consumption.
The GB200 superchip, combining CPU and GPU, outperforms its predecessors by a factor of 30, while using less power. NVIDIA's increased orders with TSMC by 25% reflect the high demand.
Blackwell is set to become the standard for AI training and various applications, from robotics to healthcare. The Supermicro version of the GB200 NVL72 showcases a compact, efficient design, using liquid cooling to manage its substantial power needs.
In essence, NVIDIA's investment in Blackwell is a bold move in the AI race, promising high returns despite the initial costs.
| Scores | Value | Explanation |
|---|---|---|
| Objectivity | 6 | Comprehensive reporting and in-depth analysis. |
| Social Impact | 5 | Significantly influencing public opinion in AI. |
| Credibility | 5 | Solid evidence from authoritative sources. |
| Potential | 6 | Inevitably leading to significant changes. |
| Practicality | 5 | Widely applied in practice, good results. |
| Entertainment Value | 2 | Includes a few entertaining elements. |