Intel Habana Gaudi Beats Nvidia H100 in AI Race

Intel Habana, a specialized division of Intel focusing on artificial intelligence accelerators, has announced that its Gaudi chips have demonstrated superior performance compared to NVIDIA competitors in artificial intelligence models for visual and linguistic analysis (Vision-Language, VL) developed by Hugging Face, a renowned company in the field of natural language and computer vision.

Hugging Face is widely recognized for its development and research efforts in the areas of natural language processing and computer vision. The company has created several popular AI models, including BERT, GPT-2, and GPT-3, which possess capabilities such as text generation, emotion analysis, question-answering, and more. Among these models, Bridgetower stands out for its ability to process and establish connections between language and visual information.

Intel Habana attributes the significant performance enhancement to the hardware acceleration of data loading, which is often a bottleneck in training AI models, especially VL models. Loading the working load into memory can pose challenges for calculations, making it a critical area of optimization for Intel Habana. The primary issue arises from the CPU being burdened with resource-intensive tasks like image decoding and augmentation, similar to the GPU’s rendering problem. As a result, the HPU (or NVIDIA GPU) has to wait for the data to be processed by the TsPU before it can be sent to the accelerator.

In tests conducted by Intel Habana using a pre-trained Bridgetower model with 866 million parameters, it was revealed that the Gaudi chips outperform NVIDIA A100 80 GB chips by a factor of 2.5. Moreover, the Gaudi chips even surpass the state-of-the-art NVIDIA H100 chip by 1.4 times.