举世瞩目的英伟达 GTC 2024 大会已拉开序幕。周一的开场演讲中，英伟达的下一代 AI 芯片，即 H100 芯片的继任者没有被命名为 B100，而是被称作 Blackwell。该芯片比上一代速度更快、也更大，正如黄仁勋在台上演示时说：“我们需要更大的 GPU，因此，女士们、先生们，我要向你们介绍一款非常大的 GPU。”Blackwell 芯片将于今年晚些时候上市。瑞银分析师估计，英伟达的新芯片价格可能高达 5 万美元，约为上一代芯片价格的两倍。黄仁勋说，新的芯片可以在大约 90 天内使用 2,000 个 Blackwell GPU 训练最新的大型 AI 模型，耗电量仅为 4 兆瓦。若使用旧款芯片，这一过程将需要 8,000 个 GPU，耗电 15 兆瓦，耗时相同。在技术性极强的演示中，黄仁勋举例说明了开发和使用 AI 系统的客户可以通过哪些不同方式使用这些新芯片。黄仁勋还介绍了为 AI 应用设计的新型网络交换机和集成了这些新芯片的超级计算机系统。黄仁勋还谈到了他对一些更遥远技术的愿景，包括药物发现、基因组学和人型机器人。
The highly anticipated NVIDIA GTC 2024 conference has kicked off. In Monday’s keynote address, NVIDIA’s next-generation AI chip, the successor to the H100 chip, was revealed and named Blackwell instead of B100. This chip is faster and larger than its predecessor, as demonstrated by Jensen Huang on stage, who said, “We needed a bigger GPU, so ladies and gentlemen, I present to you a very big GPU.” The Blackwell chip is set to be released later this year. Analysts at UBS estimate that NVIDIA’s new chip could be priced as high as $50,000, which is roughly double the price of the previous generation chip. Huang stated that the new chip can train the latest large-scale AI models using 2,000 Blackwell GPUs within approximately 90 days, consuming only 4 megawatts of power. The same process using the older chip would require 8,000 GPUs, consuming 15 megawatts of power, while taking the same amount of time. In a highly technical demonstration, Huang provided examples of how customers developing and utilizing AI systems can leverage these new chips in various ways. He also introduced new network switches designed for AI applications and supercomputing systems integrated with these new chips. Huang also discussed his vision for more distant technologies, including drug discovery, genomics, and humanoid robots.
微信扫一扫