The race to develop powerful, efficient artificial intelligence (AI) hardware took a significant step forward this week, with Intel and Google both revealing new chips aimed at reducing their dependence on NVIDIA’s technology. While NVIDIA’s GPUs currently power a majority of cloud computing data centers used for AI model training, both tech giants are looking to shift the balance.
At the forefront of Intel’s announcement is their new AI accelerator chip, Gaudi 3. In comparison with NVIDIA’s H100 GPUs that have been the standard in AI data centers, Intel claims that Gaudi 3 offers 50% better inference and 40% better power efficiency, all at a lower cost. This increased power efficiency is largely due to Intel’s use of Taiwan Semiconductor Manufacturing Co’s 5nm process in manufacturing the chips, which leads to power savings. While Intel has not revealed any specific pricing, they are confident that their offering will be “highly competitive” next to NVIDIA’s products. Dell Technologies, Hewlett Packard Enterprise, Lenovo, and Supermicro are slated to be the first to integrate the Gaudi 3 into their AI data centers.
On the other hand, Google’s focus lies in its newly announced custom Arm-based CPUs called Axion, designed to power its data centers. Compared to the fastest Arm-based instances available in the Cloud, it promises a 30% better performance and up to 60% better energy efficiency than comparable current-generation x86-based instances. Google’s move to incorporate Axion chips is strategic, making it easier for customers to move their CPU-based AI applications to Google’s cloud platform without needing significant redesigns.
Moreover, Google has announced upgrades to its Tensor Processing Unit (TPU) chips, used as NVIDIA’s alternatives for large-scale model training. The latest TPU v5p contains twice the number of chips found in the current TPU v4 pod.
However, unlike Intel, Google is not seeking to sell these chips. Instead, their focus is on drawing users to their cloud computing services rather than directly compete with NVIDIA as a hardware supplier.
These chip upgrades are also set to contribute to Google’s ‘AI Hypercomputer’, facilitating large-scale AI model training. Presently, the AI Hypercomputer also uses NVIDIA H100 GPUs, but Google plans to replace these soon with NVIDIA’s new Blackwell GPUs.
These latest developments in AI hardware indicate that the increasing demand for AI chips is getting diversified. NVIDIA’s dominant position in this arena is being challenged, with Google and Intel unveiling strong contenders to compete with NVIDIA’s offerings.