Benchmark results released by Intel suggest that Gaudi 2 accelerators are capable of outperforming NVIDIA GPUs in artificial intelligence. According to tests carried out by Stability AI, the Gaudi 2 were 55% faster than NVIDIA H100 and three times faster than the A100 in synthetic AI benchmarks.
This comparison is made considering only NVIDIA solutions currently available on the market, as Blackwell architecture accelerators will be introduced next week during GTC 2024. Intel is also developing Gaudi 3 accelerators to compete with the next generation of NVIDIA GPUs for AI.
The tests involved the Stable Diffusion 3 model, used for text-to-image generative AI. The Gaudi 2 achieved 927 images per second, while the NVIDIA H100 reached 595 and the A100 only 381.
While NVIDIA GPUs largely dominate today's AI infrastructure, especially in model-based services like GPT, Intel is rapidly optimizing the performance of its accelerators to establish a competitive offering in the market.
Intel's goal is not to force companies that already use AI accelerators to migrate to Gaudi, but rather to offer a competitive alternative. This can secure contracts with customers who have not yet adopted AI platforms, as well as enable infrastructure upgrades and expansions.
It's important to note that GPUs are more efficient at training natural language models, while inference runs are typically processed on CPUs.