Tachyum’s first chip Prodigy has not completely finished its design but one customer will buy hundreds of thousands of Universal chips to build a 50 exaFLOPS supercomputer.
Tachyum® announced that it has accepted a major purchase order from a US company to build a large-scale system, based on its 5nm Prodigy® Universal Processor chip, which delivers more than 50 exaflops performance that will exponentially exceed the computational capabilities of the fastest inference or generative AI supercomputers available anywhere in the world today. When complete, the Prodigy-powered system will deliver a 25x multiplier vs. the world’s fastest conventional supercomputer built just this year, and will achieve AI capabilities 25,000x larger than models for ChatGPT4.
Tachyum has developed the world’s first Universal Processor that combines the functions of a CPU, GPU, and TPU into a single homogeneous processor architecture that is faster, 10x lower power, and 1/3 the cost of competing products.
The Tachyum solution delivers never-before seen performance and efficiency to a wide range of applications including Hyperscale, HPC, and AI.
The Prodigy system with hundreds of petabytes of DRAM will have 100x more memory than needed to have human brain level compute.
Installation of this Prodigy-enabled solution will begin in 2024 and reach full capacity in 2025. Among the deliverables are:
* 8 Zettaflops AI training for big language models
* 16 Zettaflops of image and video processing
* Ability to fit more than 100,000x PALM2 530B parameter models OR 25,000x ChatGPT4 1.7T parameter models with base memory and 100,000x ChatGPT4 with 4x of base DRAM
* Upgradable memory of the base model system
* Hundreds of petabytes of DRAM and exabytes of flash-based primary storage
* 4-socket, liquid-cooled nodes connected to 400G RoCE ethernet, with the capability to double to an 800G all non-blocking and non-overprovisioned switching fabric
Tachyum’s proprietary TPU® AI Inference IP supports Tachyum AI (TAI) data type and provides even more; breakthrough efficiency for video and large language model data formats that otherwise would require excessive power and expensive multipliers in matrix multiplication to achieve.
As a Universal Processor offering utility for all workloads, Prodigy-powered data center servers can seamlessly and dynamically switch between computational domains (such as AI/ML, HPC, and cloud) on a single architecture. By eliminating the need for expensive dedicated AI hardware and dramatically increasing server utilization, Prodigy reduces CAPEX and OPEX significantly while delivering unprecedented data center performance, power, and economics. Prodigy integrates 192 high-performance custom-designed 64-bit compute cores, to deliver up to 4,5x the performance of the highest-performing x86 processors for cloud workloads, up to 3x that of the highest performing GPU for HPC, and 6x for AI applications.
Tachyum is transforming the economics of AI, HPC, public and private cloud workloads with Prodigy, the world’s first Universal Processor. Prodigy unifies the functionality of a CPU, a GPGPU, and a TPU in a single processor that delivers industry-leading performance, cost, and power efficiency for both specialty and general-purpose computing. When hyperscale data centers are provisioned with Prodigy, all AI, HPC, and general-purpose applications can run on the same infrastructure, saving companies billions of dollars in hardware, footprint, and operational expenses. As global data center emissions contribute to a changing climate, and consume more than four percent of the world’s electricity—projected to be 10 percent by 2030—the ultra-low power Prodigy Universal Processor is a potential breakthrough for satisfying the world’s appetite for computing at a lower environmental cost. Prodigy, now in its final stages of testing and integration before volume manufacturing, is being adopted in prototype form by a rapidly growing customer base, and robust purchase orders signal a likely IPO in late 2024.
The design will deliver 20 ExaFlops of FP64 vector performance in a 60MW power envelope and footprint of 6,000 square feet. This meets the requirements of the US Department of Energy for supercomputing systems for both HPC and AI.
The design was developed by Tachyum’s systems, solutions, and software engineering teams to allow HPC and AI workloads to run on the same architecture rather than separately on CPUs and GPUs. This will use the 3nm Prodigy 2 chip which expected to be available in 2025 with more than the 128 64bit processing cores in the first generation.
This same data centre can deliver over 10 ZetaFlops of AI performance, more than 30x the top end of the US target says Tachyum with FP64 performance up to 25 EF within the same 6,000 square-foot area, depending on the flexibility of the power envelope.
Prodigy has custom 48U rack reference designs that come in two versions, a 1U liquid-cooled version and a 2U air-cooled card. The liquid-cooled version supports 88 4-socket 1U servers for a total of 352 Prodigy Processors, and the air-cooled version supports 40 4-socket 2U servers for a total 160 Prodigy Processors. The racks have a modular architecture, and 2 racks can be combined into a 2-rack cabinet to optimize floor space.
The Tachyum Prodigy 2 processor was selected by Important Project of Common European Interests (IPCEI) program for Slovakia to deliver exa-scale HPC and zetta-scale AI for Europe. The European Commission has accepted the funding gap of €26.4m for Tachyum, which is currently in the notification process.
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.
Currently, the US has a 1.68 (Frontier at Oakridge) and the Chinese record might (they don’t talk about it much) be 1.72 (both measured at FP64 precision). But, regardless of peak, the Chinese machine has only demonstrated a sustained performance at 115.8 petaflops (about 12% of an exaflop). Meanwhile, the Lawrence Livermore National Laboratory is promising its El Capitan at 2 exaflops. It’s barely possible the Chinese might be working on a 3 exaflop machine, but there are a lot (and I do mean a lot) of “ifs” in that conjecture.
In the light of all this, a 50 exaflop machine showing up next year, not to mention especially designed for AI, would be bigger than huge, it would be positively disruptive–especially if they can deliver full capacity by 2025, as that would be skipping us about a full decade into the projected future in a single instant. On the face of it, that seems a pretty big “if” but I’ll be keeping my fingers crossed.
100*the ability of a human brain? Hope it comes fitted with an off switch.
As a non-expert who has always had an interest in technology and how it affects society, this latest AI wave has been fascinating. When the LLMs first started hitting the media I thought it was a bit overblown to call them AI. Granted they are more complex, powerful and useful than scripted chat it’s of the past (or NPCs in games which were also referred to casually as “AI”) but I considered them to be “simulated” intelligence. Frankly I don’t believe in “narrow” AI. Once something is close enough to real intelligence to not be considered simulated it would qualify as AGI.
Now though, the transition to large multimodal models and some of the other research and development coming down the line are starting to move into a grey area after simulated intelligence but before AGI. Pieces of the puzzle from long-term memory and continuous retraining and creating their own training video synthetically which resembles imagination is kind of a big deal. I’m really starting to think this is headed in the direction of real intelligence (if it can be made into something active that doesn’t just wait for prompts.
The point I’m finally getting to is that compute demand is going to be out stripping supply massively in the near term.
AI is a very overused term and has become rather vague, covering many very different sorts of things. Most of these won’t be beings like us. They will be more like genies. They will wait placidly, doing whatever they are supposed to be doing, until someone makes a wish. Then they try to make it come true. We must be very careful in how we word our wishes and, especially in the case of the more powerful genies, who is allowed to request wishes from them.
I believe our world is already in an arms race to achieve these devices, these artilects. It’s quite possibly going to be winner take all because the first country that succeeds may be able to monitor everything, run everything, control everything, and enforce anything. (Kind of what China is attempting to do right now.) Thereby being in a position to effectively end any other organization’s attempt to achieve their own versions of advanced AI.
No one can drop out of this race because, if they don’t build it, someone else will build it (to use an old but still valid argument). To paraphrase science fiction writer Larry Niven, “What a terrible thing if any faction other than my own should gain control of this!”