近日，由Stable Diffusion开发者Stability AI发布的一份报告显示，英特尔的Gaudi 2计算卡在人工智能图像生成训练方面，相比英伟达的顶级H100芯片，展现出显著的效率优势。据该报告详细阐述，以2亿参数的MMDiT模型为基准，训练深度设定为24，并采用BFloat16的混合精度计算，Gaudi 2芯片在最高配置下能实现每秒训练1254张图片，而在256容量配置下则可达到每秒927张的训练速率。
相比之下，英伟达的H100-80GB计算卡在同一模型设置下，每秒仅能训练595张图片，而A100-80GB芯片的训练速度更是降至每秒381张。这意味着在相同容量下，Gaudi 2的训练速度比H100提升了55%，其性能是A100的2.43倍。这一结果无疑标志着英特尔在AI计算领域的重大突破，为高性能计算和大规模AI应用提供了更高效的选择。
此次报告的发布，揭示了英特尔Gaudi 2芯片在AI训练效率上的显著提升，可能对未来的AI开发和数据中心市场产生深远影响。随着技术的不断进步，两家公司在AI计算领域的竞争也将更加激烈。
英语如下：
**News Title:** “Intel’s Gaudi 2 Chip Makes a Splash: Outperforms NVIDIA in AI Image Generation Training by 55%”
**Keywords:** Intel Gaudi 2, AI image generation training, performance boost
**News Content:**
In a recent report by Stable Diffusion developer Stability AI, Intel’s Gaudi 2 accelerator card has demonstrated superior efficiency in artificial intelligence image generation training compared to NVIDIA’s top-of-the-line H100 chip. The report details that, using a 200 million parameter MMDiT model as the baseline, with a depth setting of 24 and BFloat16 mixed-precision computations, the Gaudi 2 chip can achieve a training rate of 1,254 images per second at its peak configuration, and 927 images per second at a 256 capacity setup.
In contrast, NVIDIA’s H100-80GB card manages only 595 images per second under the same model settings, while the A100-80GB chip drops to 381 images per second. This indicates a 55% training speed improvement for Gaudi 2 over the H100 at equivalent capacity, and a 2.43 times performance increase compared to the A100. This milestone underscores Intel’s significant breakthrough in AI computing, offering a more efficient option for high-performance computing and large-scale AI applications.
The revelation of Gaudi 2’s enhanced AI training efficiency in the report is expected to have far-reaching implications on future AI development and the data center market. With ongoing technological advancements, the competition between Intel and NVIDIA in AI computing is poised to intensify.
【来源】https://www.itheat.com/view/45945.html
Views: 1