Future Tech

Intel woos China with nerfed Habana Gaudi 2 AI chips

Tan KW
Publish date: Thu, 13 Jul 2023, 10:30 AM
Tan KW
0 461,939
Future Tech

Intel has followed Nvidia's lead and will produce a modified version of its AI accelerator - specifically the Habana division's Guadi 2 - for the Chinese market.

Chen Baoli, VP and GM of Intel's Datacenter and AI Group in China, is understood to have announced availability of the chips during a press conference in Beijing this week. Several Chinese server vendors - including Inspur, H3C, and xFusion - are said to have lined up to sell them.

Introduced in spring 2022, Gaudi 2 was designed to compete with Nvidia's A100 GPUs, which have been a popular choice for enterprises training large language models (LLMs). For those who don't recall, Intel bought Habana for $2 billion in late 2019. The Gaudi 2 processor itself is fabricated using a 7nm process node and features 24 tensor cores, 96GB of HBM2e memory, and 24 100GbE ports.

At least according to Intel's internal benchmarks, the accelerator boasted roughly twice the performance in the ResNet-50 image classification and BERT natural language processing model tests compared to Nvidia's A100. Despite this, Gaudi 2 hasn't enjoyed the widespread adoption that Nvidia's GPUs have. It probably didn't help that Nvidia's H100, announced a few months earlier, promised performance 6x that of the A100.

Further complicating matters, US regulation on the export of AI accelerators, implemented in November, has made it difficult for Chinese companies to get their hands on GPUs and other chips used for LLMs and other machine learning applications.

Under the rules, chipmakers are barred from selling processors capable of greater than 600GB/sec of I/O bandwidth in China. The limit impacted many of the more popular GPUs used in AI and high-performance computing applications - including Nvidia's A100 and H100 as well as AMD's Instinct MI250X.

In the wake of the decision, Nvidia released a slower version of its A100, called the A800, with half the memory and about two thirds the bandwidth of the full-fat card. Since then, the acceleration specialist has also announced a China-safe version of its H100 - predictably enough dubbed the H800.

According to DigiTimes today, Intel's Gaudi 2 will also need to be nerfed to a degree before it can be sold in the Chinese market. However, it's not clear how the chip will need to change. As we understand it, the standard chip should already squeak by under the I/O limit. However, the US government has recently signaled that it may tighten restrictions on AI chip exports even further.

The Register has reached out to Intel for comment and we'll let you know if we hear anything back.

Having said that, this wouldn't be the first time Intel has reworked one of its chips for "other markets." In April we reported that Intel planned to launch a GPU Max SKU, called the 1450, with lower I/O bandwidth and support for either air- or liquid-cooled systems. The reduction in I/O bandwidth led us to believe Intel had the Middle Kingdom in its sights, though the chip shop stopped short of naming China as the target market. ®

 

https://www.theregister.com//2023/07/13/intel_guadi_china/

Discussions
Be the first to like this. Showing 0 of 0 comments

Post a Comment