Tachyum, a Slovak/US company, has revealed its groundbreaking supercomputer, expected to be operational by 2025, boasting an astonishing 50 exaFLOP performance, setting it 25 times ahead of contemporary systems. The key to this colossal leap in computing power lies in Tachyum’s Prodigy chip, a technological marvel that promises to redefine the landscape of data centers, AI, and high-performance computing across various sectors.
Tachyum’s prodigy chip: Revolutionizing computing architectures
The Prodigy chip, at the heart of this revolutionary supercomputer, represents a significant milestone in memory, storage, and compute architecture. Designed to cater to the diverse workloads of government, research, academia, business, manufacturing, and more, the Prodigy chip is set to unlock unprecedented capabilities.
Earlier this year, Tachyum provided a glimpse of its ambitions by unveiling a supercomputer architecture based on the Prodigy chip, targeting 20 exaflops of computing power. With plans for sampling the chip next year, Tachyum’s vision is steadily becoming a reality.
The installation of this Prodigy-enabled supercomputer is scheduled to commence in 2024, with full capacity expected to be reached in 2025. Once operational, it will deliver a staggering 8 Zettaflops of AI training capacity for large language models, along with an additional 16 Zettaflops dedicated to image and video processing.
This level of computational prowess equips the supercomputer to accommodate more than 100,000 times the PALM2 530B parameter models or 25,000 times the ChatGPT4 1.7T parameter models with base memory. Furthermore, it can support 100,000 ChatGPT4 models with four times the base DRAM.
The infrastructure of the Prodigy-enabled supercomputer is as impressive as its computational capabilities. The system comprises 4-socket, liquid-cooled nodes connected through a 400G RoCE Ethernet network. Notably, this network has the potential to double its capacity to an astounding 800G, all while maintaining non-blocking and non-overprovisioned switching capabilities.
Tachyum’s visionary leap
Dr. Radoslav Danilak, the founder and CEO of Tachyum, underscores the remarkable achievement of the Prodigy chip and the supercomputer it powers. He states, “The unprecedented scale and computational power required as part of this installation simply could not be provided by any chip manufacturer on the market today.”
Dr. Danilak further emphasizes that while some startups have garnered substantial investments for their promises of achieving similar capabilities in the future, only Tachyum is poised to deliver economically viable, order-of-magnitude larger machines that could pave the way for the transition to cognitive AI.
Central to Tachyum’s game-changing approach is its Universal Processor. This processor is designed to seamlessly handle a wide range of workloads, eliminating the need for dedicated AI hardware. Servers equipped with the Universal Processor can dynamically switch between computational domains, such as AI/ML, HPC, and cloud computing, all on a single architecture. This not only increases server utilization but also significantly reduces capital and operational expenses.
