UAE supercomputer powers up to fight global rivals

Mohamed bin Zayed University of Artificial Intelligence

• New machine is Condor Galaxy 1
• Used to train Jais language model
• $100m deal to provide computers to G42

A supercomputer built in the US with UAE backers is set to challenge global industry artificial intelligence chip giants after its success powering a new advanced large language model (LLM).

The 13 billion-parameter Arabic and English LLM, known as Jais, was unveiled on Wednesday.

Local industry experts said it appears to be more efficient and cost-effective than global rivals, and its developers told AGBI they plan to adapt it to other languages and other media such as voice, imagery and video.

• US and UAE supercomputer venture advances AI training
• Marketers are grappling with AI’s role in advertising
• UAE to play ‘central role’ in AI but needs greater regulation

The model is a collaboration between Abu Dhabi’s Mohamed bin Zayed University of Artificial Intelligence (AI), Silicon Valley-based Cerebras Systems and G42, an AI company chaired by the UAE’s national security adviser Sheikh Tahnoon bin Zayed al Nahyan.

Jais was trained on supercomputer Condor Galaxy 1 (CG1). According to Cerebras, CG1 can support up to 600 billion parameters and can extend up to 100 trillion (a parameter is the value used to configure a learning algorithm). This makes it one of the largest AI supercomputers in production.

It is difficult to compare different processors involved in big LLM projects, since the high-performance computing systems involved are highly customised.

However, UAE-based technology consultant Carrington Malin said that CG1’s performance could give major chipmakers a run for their money. 

“It’s common for LLMs to take months to train, but Jais was trained in just 21 days,” he told AGBI.

“CG1 is a super high-performance computer that is providing the compute capacity needed at a fraction of the cost of many alternatives.” 

Gulf states are expected to invest heavily in powerful semiconductor chips in the years ahead. The UAE and Saudi Arabia have reportedly been buying up thousands of Nvidia chips needed to drive their AI ambitions.

The computing power provided by supercomputers is vital for building AI products such as the advanced LLM chatbot ChatGPT.

US chipmaker Nvidia – the world’s most valuable manufacturer, worth nearly $1 trillion – holds a 95 percent market share.

It dominates the market for high-performance chips known as graphics processing units (GPUs) that process vast amounts of data much faster than traditional computer components.

While Nvidia’s US rival AMD and other chip manufacturers have made significant strides in GPU technology, its GPUs have remained the preferred choice for many developers and organisations.

G42 is backed by private equity giant Silver Lake. Andrew Jackson, CEO of Inception, the G42 subsidiary that developed Jais, told AGBI the company is “doubling down on Cerebras” after using CG1.

He said it achieves the performance of larger models that are far more expensive to run in production.

Earlier this year, Cerebras signed a $100 million deal to deliver the first of what could be up to nine AI supercomputers to G42.

“We are massive fans of Nvidia and we have been their largest customer in the region,” Jackson said. 

“But we’re also very open to trying new things. We have run on other hardware and when we got access to CG1 we saw a dramatic speed-up in terms of our training time.”

Inception said in a press statement that Jais outperforms existing Arabic models by a sizable margin.

It is also competitive with English models of similar size despite being trained on significantly less English data.

Jackson said CG1’s CS-2 architecture is different from that of Nvidia’s GPUs – and the former enabled a custom-built machine for AI training.

“It’s focused on specific tasks and is less generalised than Nvidia hardware, which basically means for us it’s more dedicated for solving AI training problems,” he said. 

Reuters

The surge of interest in AI has contributed to a huge supply shortage of Nvidia’s semiconductors, resulting in long waiting lists, and driving companies around the world to seek alternatives.

“We are blessed to have a very large repository of compute but if you speak to any AI scientist, they’ll always want more,” Jackson said.

“We all have backlogs of things we want to do and try. If I had more compute today, I would run a lot more experiments.”

Named after the highest mountain in the UAE, Jais will be available via an open-source licence.

By open-sourcing Jais, Inception aims to engage scientific, academic and developer communities to accelerate the growth of an Arabic-language AI ecosystem for the world’s 400 million-plus Arabic speakers. This is currently challenged by a lack of enough Arabic data to train a model of Jais’ size.

The Jais model is being aimed at government, financial, energy, climate and healthcare domains, where it could be used to generate responses to questions, documents, translations and emails, advice, and recommendations, in Arabic and English.

There are plans to expand Jais to other languages that are under-represented in mainstream AI, as well as a multimodal model including voice and imagery and video, Jackson said. 

A report by UBS highlighted the Middle East as the next significant technology frontier market, projecting its digital economy will grow from $180 billion in 2022 to $780 billion in 2030, with annual growth of 20 percent.