Intel's New AI Chip
Intel looks to compete with NVIDIA!
Created on March 11|Last edited on March 11
Comment
Stability AI has recently shared insights into its computational benchmarks, focusing on the performance of its upcoming models: Stable Diffusion 3 and Stable Beluga 2.5. As part of their series aimed at providing a clearer view of the computational efforts behind their generative AI technologies, they compared some of their newest models on Intel and NVIDIA chips.
Clash of the 'Titans'
The performance analysis conducted by Stability AI compared Intel Gaudi 2 accelerators with Nvidia's A100 and H100 GPUs, which are widely used by startups and developers for training large language models (LLMs). The benchmarks focused on the training speeds of two of Stability AI's models, showcasing the strengths and capabilities of their computational frameworks.
Stable Diffusion 3 Insights
Stable Diffusion 3, a sophisticated text-to-image diffusion transformer, is set for an early preview release. The model, with 2 billion parameters, demonstrated that the Gaudi 2 system processed training images at a significantly higher speed compared to Nvidia's H100 and A100. Notably, the Gaudi 2's high memory capacity allowed for larger batch sizes, enhancing the training rate. While the Gaudi 2 matched Nvidia's A100 in inference speeds, it lagged behind the optimized performance of A100 with TensorRT, though there is room for optimization and improvement.
Stable Beluga 2.5 Findings
The language model Stable Beluga 2.5, an enhancement of the LLaMA 2 70B, underwent benchmarks using 256 Gaudi 2 accelerators. It exhibited outstanding performance without needing additional optimizations, surpassing the A100 with TensorRT-LLM by 25 percent in terms of tokens processed per second during both training and inference tests. The initial results are encouraging and suggest that performance could be further improved with the introduction of FP8 precision.
Stability AI highlights the growing need for more efficient and potent computing solutions, pointing out that the Gaudi 2 chips not only outperform other 7nm chips but also offer advantages like cost-effectiveness and shorter lead times. It will be very interesting to see new and existing companies innovate in this space!
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.