Cerebras introduces next-gen AI chip for GenAI training – TechTarget

author
3 minutes, 18 seconds Read
image

AI startup Cerebras Systems on Wednesday introduced the next generation of its AI chip, Wafer-Scale Engine 3.

WSE-3 delivers two times the performance of the previous generation WSE-2, according to Cerebras.

The new, super-fast chip is built for organizations looking to train large AI models and powers Cerebras’ CS-3 AI supercomputer.

Four system configurations of CS-3 can fine-tune 70B parameter models, like Meta’s Llama 2, in one day, the AI startup said.

Cerebras also revealed it has partnered with Abu-Dhabi-based technology holding group G42 to build the third cluster of its constellation of AI supercomputers, the Condor Galaxy 3. 

This is the third of nine supercomputers the vendor plans to build. Condor Galaxy 3 will deliver 8 exaflops of AI processing power with 58 million AI-optimized cores, Cerebras said. That scale of infrastructure makes the system one of the most advanced of its kind, optimized for large scientific simulations and training the biggest AI models for applications such as image and speech recognition.

Large models

The new AI chip comes as more AI vendors seek to build large models.

Despite the growing popularity of small language models, large language models will continue to advance, Tirias Research founder Jim McGregor said.

“If you think about where we are in GenAI today, it’s still very limited,” McGregor said.

The progression toward video generation will require larger compute capacity, leading to ever larger models, he said.

According to analyst firm Intersect360 Research‘s survey of 163 high-performance computing users, more than a third want to build their own generative AI models.

Cerebras’ WSE-3 AI chip and CS-3 systems could benefit those users, Intersect360 Research CEO Addison Snell said.

“They’re doubling the performance within the same cost and power,” he said. “That’s a big deal with regard to other accelerators that people are using for these types of systems.”

Advantages and disadvantages

One of Cerebras’ advantages in the market is its ability to process large amounts of data in an extremely short time, McGregor said.

However, despite the 2016 startup’s rapid growth, to the point at which it is valued at more than $4 billion, and its ability to increase the scalability of its platform, it is still a smaller company compared to the dominant AI hardware/software vendor, Nvidia, he added.

“They’re focused on one aspect of AI and that is training,” McGregor said, noting that training is only a niche in the LLM market.

Meanwhile, Nvidia has many offerings for training and other products such as its CUDA classical-quantum platform that are popular with researchers and data scientists.

However, Cerebras beknow is branching out to other areas beyond training.

Along with its WSE-3, the AI startup revealed it will use its CS-3 AI accelerators to train Qualcomm Technologies’ AI 100 Ultra system to speed up inferencing. The AI 100 Ultra is a cost-optimized inferencer for generative AI and LLMs.

The partnership gives Cerebras what some consider an inferencing product offering that is an alternative to Nvidia’s and other vendors’ systems, McGregor said.

“It’s not cost-efficient to run a workload solution on their platform,” he said.

Cerebras also has an advantage because it can process LLMs in a single system rather than having to scale thousands of GPUs, he added.

However, users of its system will have to run it all the time to realize the benefit.

“If you’re running a large language model in a day, and then you’re not running anything the rest of the week, you’re not getting a positive return on investment on that system,” McGregor said. “You need to keep those systems up and running as much as possible if you’re going to get positive return on investment.”

Despite being a startup, Cerebras is in a good position because of the current demand for generative AI, Snell said.

“The current demand for generative AI that’s outstripping supply from Nvidia and others serves to help Cerberus or any other company with [AI offerings] right now,” he said.

Esther Ajao is a TechTarget Editorial news writer and podcast host covering artificial intelligence software and systems.

This post was originally published on this site

Similar Posts