Tachyum on Tuesday mentioned that it had submitted a bid to the Division of Vitality to construct a 20 exaflops supercomputer in 2025. The machine could be primarily based on the corporate’s next-generation Prodigy processors that includes a proprietary microarchitecture that can be utilized for several types of workloads.
The U.S. DoE desires a 20 exaflops supercomputer with a 20MW–60MW energy consumption to be delivered by 2025. The system is about to be put in at Oak Ridge Nationwide Laboratory (ORNL) and can complement the lab’s Frontier system that went on-line earlier this yr.
Tachyum doesn’t disclose which {hardware} it proposed to the DoE, however solely says that it has its 128-core Prodigy processor as we speak in addition to a higher-performing Prodigy 2 processor in its roadmap, so it’s protected to say that by 2025 it’s going to have the latter readily available and it may have the ability to tackle the upcoming system.
Tachyum’s Prodigy is a common homogeneous processor packing as much as 128 proprietary 64-bit VLIW cores that characteristic two 1024-bit vector models per core and one 4096-bit matrix unit per core. Tachyum anticipated its flagship Prodigy T16128-AIX processor (opens in new tab) to supply as much as 90 FP64 teraflops for HPC in addition to as much as 12 ‘AI petaflops’ for AI inference and coaching (presumably when working INT8 or FP8 workloads). Prodigy consumes as much as 950W and makes use of liquid cooling.
That was all earlier than Tachyum sued Cadence, its mental property supplier, for lower-than-expected efficiency of its Prodigy processor. We do not know what the present efficiency expectations are for the chip.
In idea, Tachyum may energy an exaflops system utilizing over 11,000 of its Prodigy processors, although energy consumption of such a machine could be gargantuan. Presumably, Prodigy 2 has a greater probability to fulfill the wants of a next-generation exascale system than the unique Prodigy.
There may be at present one exaflops-class supercomputer within the U.S., the 1.1 exaflops Frontier system at Oak Ridge Nationwide Laboratory (ORNL) that’s primarily based on AMD’s 64-core EPYC CPUs in addition to Intuition MI250X compute GPUs. There are two extra exascale programs being constructed within the USA, the two exaflops Aurora machine powered by Intel’s 4thGeneration Xeon Scalable processors and Xe-HPC compute GPUs (aka, Ponte Vecchio) in addition to the “>2 exaflops” El Capitan supercomputer primarily based on AMD’s Zen 4 structure EPYC CPUs and Intuition MI300 GPUs.
One of many fascinating issues in regards to the DoE’s supercomputing plans is that any longer it desires to improve its high-performance compute capabilities each 12–24 months, not each 4–5 years. Consequently, the DoE shall be extra wanting to undertake unique architectures like Tachyum’s Prodigy than it’s as we speak.
“We additionally want to discover the event of an strategy that strikes away from monolithic acquisitions towards a mannequin for enabling extra fast improve cycles of deployed programs, to allow quicker innovation on {hardware} and software program,” a DoE doc reads. “One doable technique would come with elevated reuse of present infrastructure in order that the upgrades are modular. A objective could be to reimagine programs structure and an environment friendly acquisition course of that enables steady injection of technological advances to a facility (e.g., each 12–24 months somewhat than each 4–5 years). Understanding the tradeoffs of those approaches is one objective of this RFI, and we invite responses to incorporate perceived advantages and/or disadvantages of this modular improve strategy.”
One of many benefits that Tachyum’s Prodigy has over conventional CPUs and GPUs for AI and HPC workloads is that it’s tailor-made for each forms of workloads, which is why Prodigy can be utilized for AI workloads when its HPC capabilities usually are not used and vice versa. The DoE might or might not undertake Tachyum for any of its upcoming supercomputers, however the firm hopes to be awarded with an applicable contract.
