AI chip company Cerebras announces major advances in materials science, sparse training and more

Artificial intelligence chip startup Cerebras Systems Inc. says it’s getting some serious traction in its bid to rival Nvidia Corp. in powering artificial intelligence workloads.

The company announced today it has made some groundbreaking achievements in terms of molecular dynamics simulations and sparse training, advancing materials research and enabling powerful large language models to be trained more quickly at significantly lower costs.

The breakthroughs came as the company announced it has entered into a multiyear partnership with Aleph Alpha GmbH, which is regarded as one of Europe’s hottest generative AI startups. Cerebras said it will work with Aleph Alpha on the creation of “secure sovereign AI solutions” for government agencies, with its first customer being Germany’s armed forces.

Cerebras is a producer of AI-focused chip technologies that it says provide a more powerful and cost-effective alternative to Nvidia’s graphics processing units, which power the vast majority of AI workloads today. The company’s flagship product is the WSE-3 processor, which was announced in March and builds on its earlier WSE-2 chipset that debuted in 2021.

Slated to become available later this year, the WSE-3 chip is built on an advanced five-nanometer process. It features 1.4 trillion more transistors than its predecessor, with more than 900,000 compute cores and 44 gigabytes of onboard static random-access memory.

The startup will offer its WSE-3 chips, which feature 52 times more cores than a single Nvidia H100 GPU, as part of a data center appliance called the CS-3, which is about the same size as a small refrigerator. Each unit comes with a single, pizza-sized WSE-3 processor and built-in cooling and power delivery modules, which provide a standby power source in case of an outage.

Cerebras has long argued that its WSE chips are a superior alternative to Nvidia’s increasingly hard-to-come by GPUs, enhancing the efficiency of AI processing and training workloads. With today’s announcements, it’s providing compelling evidence to back up those claims.

A new standard for atomic-scale molecular dynamics simulations

The company explained that it partnered with researchers from Sandia, Lawrence Livermore and Los Alamos national laboratories, setting up a system based on its older WSE-2 chips that was able to perform atomic-scale materials simulations at the millisecond scale for the first time.

Cerebras published a research paper that shows how its simulations were performed an astonishing 179 times faster than what has been achieved by the world’s No. 1 supercomputer, Frontier, which is powered by 39,000 GPUs. “No number of GPUs can get this result,” Andy Hock, senior vice president of product and strategy, told SiliconANGLE. “We’re unlocking fundamental new timescales for molecular dynamics.”

The company explained that Frontier and other leading supercomputer platforms have always been limited to simulating materials at the atomic scale at a rate of about two to three microseconds per month. The comparative sluggishness of those platforms seriously restricts the understanding of how materials evolve and behave over long periods of time. So the ability to simulate them at exponentially greater speeds will have some serious implications for materials science, the company promised.

The feat was achieved by using a novel mapping scheme that assigns each atom to a single core on the WSE-2 chip. Those cores were then organized as a 2D grid, with the physical simulation domain mapped onto it to preserve locality. The cores then communicate with one another, exchanging atom information, enabling more efficient parallel processing.

The breakthrough promises to transform materials science research by helping scientists to study much longer-term behavior of complex materials, such as the evolution of grain boundaries in metals, potentially paving the way for the development of stronger and more durable materials. It will also help to accelerate protein folding simulations to aid in drug development, and optimize catalytic reactions to help design more efficient energy renewable energy storage systems, the company said.

“These results open up new opportunities for materials research and science discoveries beyond what we envisioned,” said Siva Rajamanickam, principal member of technical staff at Sandia National Laboratories.

Accelerated sparse training

The progress in materials science simulations demonstrates Cerebras’ prowess in old-school AI, but its CS-3 systems are also just as useful for generative AI, the company said today. To prove that, it announced the results of a long-term collaboration with Neuralmagic Inc., which develops high-performance servers for AI inference, showing how its platform was able to accelerate the training of sparse neural networks, which are modeled on the human brain.

The company said it was able to achieve an unprecedented 70% parameter reduction when training large language models on the CS-3 system and deploying them on Neural Magic’s inference server systems. With this, it says it will enable companies to design and train more efficient, lower-cost LLMs with a wide range of potential applications.

“For the first time ever, we achieved up to 70% sparsity for a foundational model, such as Llama, with full accuracy recovery for challenging downstream tasks,” said Cerebras co-founder and Chief Technology Officer Sean Lie.

Cerebras and Aleph Alpha to partner on secure sovereign AI

Building on these announcements, Cerebras said, it’s partnering with the generative AI startup Aleph Alpha to train and deploy cutting-edge AI modes for the German Armed Forces and other customers.

The partnership is quite the coup for Cerebras, as Aleph Alpha becomes one of the most significant players in the generative AI industry to back the company publicly. Aleph Alpha is a rival to OpenAI that made headlines in November when it announced a $500 million-plus Series B funding round led by Hewlett Packard Enterprise Co. and SAP SE. HPE is also one of its main customers, with Aleph Alpha providing ready-to-use LLMs for the HPE GreenLake for Large Language Models cloud service.

Aleph Alpha is best known for its Luminous generative AI models, which can process text in English, French, German, Italian and Spanish, as well as images and video. The startup says its general-purposes LLMs are suitable for a wide range of tasks, such as text generation, document summaries and sorting, object classification and more. It also stands out for its AI model explainability capabilities, which help customers ensure the accuracy of its models’ responses.

It’s this capability that will enable Cerebras and Aleph Alpha to develop a new class of compute-efficient AI models that are sovereign and trustworthy, suitable for use by government agencies, said Aleph Alpha CEO Jonas Andrulis. “We chose Cerebras because of their world-class AI expertise and peerless wafer-scale technology that enables us to train state-of-the-art AI models with high efficiency,” he explained.

The startup will become the first European organization to deploy the Cerebras CS-3 platform at its alpha ONE data center, representing another milestone for the chop maker. Aleph Alpha said it will use CS-3 to lead its research efforts that are focused on the creation of new foundational models, multimodal models and architectures.

Holger Mueller of Constellation Research Inc. said it was only a matter of time before a true rival to Nvidia’s AI chips emerged, and he believes Cerebras could well be it. “It’s hardware is making number of AI workloads much faster and more affordable, and today’s partnership announcements demonstrate the viability of its offerings,” Mueller said. “While all are noteworthy, the partnership with Aleph Alpha stands out as it will enable the sovereign AI workloads that are essential for any organization doing business in Europe, which is developing strict regulations to govern AI.”

“With Aleph Alpha’s impressive track record of creating sovereign and secure AI solutions, we believe this new partnership will produce a new class of AI model architectures that governments and enterprises worldwide can benefit from, in a safe and secure way,” said Cerebras CEO Andrew Feldman.

With reporting from Robert Hof

Photo: Cerebras Systems