Cerebras and Qualcomm Drive the Future of Generative AI With New Technologies
In a significant leap forward for generative AI capabilities, Cerebras has unveiled a suite of innovations that promise to redefine the landscape of artificial intelligence computing. Among these advancements are the introduction of the Wafer Scale Engine AI processor (WSE-3), the CS-3 server, and the highly anticipated Galaxy Condor 3 (CG-3) supercomputer. Furthermore, a strategic collaboration with Qualcomm aims to enhance inference processing, marking a pivotal moment in the AI domain.
Cerebras has evolved from a mere systems vendor to a formidable service provider, joining forces with G42 to materialize plans for expanding its supercomputing centers across the United States. This shift signifies Cerebras’ transition into a prime contender in AI training, diverging from the conventional business model adopted by most AI startups. This strategic direction not only underscores the company’s technical acumen but also aligns with broader industry success narratives similar to those of semiconductor innovators like Ampere.
The technical mastery showcased by Cerebras delineates a clear distinction in the AI landscape. Through overcoming significant engineering hurdles, the company has successfully pioneered the Wafer Scale Engine (WSE), deploying a chip as large as a 200mm silicon wafer. This breakthrough has facilitated the training of colossal language models with unparalleled efficiency and accuracy. Initially gaining traction within sectors like pharmaceutical research due to the unique challenges they present, Cerebras now boasts a diversified portfolio of clientele across healthcare, energy, and more, highlighting its wide-reaching applicability.
Addressing another critical engineering challenge, Cerebras’ ability to scale its platform for data center applications has culminated in the development of the CS-2, marking the inception of its supercomputing era with the Condor Galaxy (CG) series. The launch of CG-1 and CG-2 supercomputers, engineered in partnership with G42, achieved four exaFLOPS of AI compute performance, a testament to the company’s innovative prowess.
The unveiling of the third-generation WSE-3 processor ushers in a new era of semiconductor excellence. Fabricated with TSMC’s 5nm process, this behemoth hosts four trillion transistors and 900,000 processing cores, delivering a staggering 125 petaFLOPS of AI performance. The CS-3 server, designed to house this powerful chip, exemplifies Moore’s Law, offering double the performance of its predecessor at an equivalent cost and energy consumption.
These groundbreaking advancements facilitate the efficient training of expansive Generative AI Language Models. Cerebras asserts that a single CS-3 unit can expedite the training of trillion-parameter models, significantly reducing both time and complexity involved, thereby enhancing computational efficiency and cost-effectiveness.
Simultaneously, the announcement of the Condor Galaxy 3 (CG-3) supercomputer construction in Dallas, Texas, reflects Cerebras’ commitment to scalability and performance, aiming to deliver 8 exaFLOPS of AI power by Q2 2024. This venture is part of a more extensive plan to erect nine supercomputing data centers by the end of 2024, showcasing an ambitious roadmap for the company’s expansion.
In a strategic move to bridge the gap from training to inference, Cerebras partners with Qualcomm. Leveraging Qualcomm’s Cloud AI 100 and Cloud AI 100 Ultra platforms, this partnership focuses on enhancing the efficiency of generative AI inference processing. By implementing advanced techniques such as sparsity and model optimization, the collaboration between Cerebras and Qualcomm promises to yield up to a tenfold increase in tokens per dollar, heralding a new age of inference efficiency.
Unlike their integration into the Condor Galaxy data centers, Qualcomm’s inference solutions can be deployed across various data centers, ensuring versatility and accessibility for AI and generative AI model processing.
As the AI era unfolds, Cerebras’ contributions underscore a monumental shift in the industry’s approach to addressing the burgeoning demand for powerful, efficient AI computing solutions. This narrative heralds a welcome influx of innovation and diversity, as companies like Cerebras ascend to join the ranks of semiconductor and systems leaders, paving the way for more advanced problem-solving capabilities in the realm of artificial intelligence.
