News

[News] Intel Steps Back from AI Race with NVIDIA, Shifts Focus to Cost-Effective Solutions with Gaudi 3


2024-10-18 Semiconductors editor

According to a report from Wccftech, Intel has stepped away from competing with NVIDIA in AI computing power and the market of training large-scale AI models. Instead, the company is now entering a less saturated segment of the AI market, focusing on its new cost-effective AI accelerator, Gaudi 3.

Intel aims to present Gaudi 3 as the product with the best price-to-performance ratio, though Gaudi 3 is “not catching up” to NVIDIA’s latest GPU from a head-to-head performance standpoint, citing the words from Anil Nanduri, head of Intel’s AI acceleration office, during an interview with CRN.

Nevertheless, Nanduri highlights that the Gaudi 3 accelerator chip is ideal for supporting cost-effective systems that run task-based and open-source models for enterprises.

On the other hand, according to the report from Wccftech, Intel claims that following the introduction of “reasoning-focused” LLM models, its Gaudi 3 lineup delivers performance comparable to NVIDIA’s well-known H100 AI accelerator, especially in inference workloads.

According to the report from CRN, Intel claims that Gaudi 3 is about 9% faster than H100 in the Llama 3 model and 80% more cost-effective; in the Llama 2 model, Gaudi 3 is 19% faster and the price-performance difference is up to 2 times.

However, when evaluated in terms of floating-point operations, the Gaudi 3 AI GPUs fall short compared to NVIDIA’s options, indicating that high-end AI performance isn’t currently Intel’s strength.

Therefore, according to the report from Wccftech, Intel has no plans to compete directly with NVIDIA’s GPUs. Furthermore, there is another rising competitor in the AI computing sector: AMD.

Regarding the reason behind Intel’s choice, the company believes that smaller LLM models will see increased acceptance as the initial excitement around AI and the buzz surrounding large-scale data centers diminish, as indicated by the report in Wccftech.

Read more

(Photo credit: Intel)

Please note that this article cites information from Wccftech and CRN.

Get in touch with us