Tachyum introduced the Prodigy ATX Platform, a workstation tailored for artificial intelligence applications, based on a customized version of the company’s primary processor, which is yet to be finalized, in an exclusive whitepaper release on Friday. The platform aims to democratize access to extensive language models containing billions or even trillions of parameters, with a price tag set at $5000. The exact availability date of the product remains undisclosed at this time.
The foundational element of the impressive ATX Platform is the 96-core Prodigy processor, projected to enter mass production on TSMC’s 5nm cluster by 2024, boasting a peak operating frequency of 5.70 GHz. Tachyum highlights that the CPU operates with only half of its die activated, a design choice that contributes to cost reduction, faster delivery, and decreased power consumption.
In the realm of processor development, it is noteworthy that the Prodigy’s launch, initially slated for 2020 post a 2019 tape-out, has faced multiple delays to 2021, 2022, and subsequently to 2023 despite ongoing claims of enhanced performance capabilities. The latest update from the company indicates a targeted release for the second half of 2024, possibly in December, without elaborating on the reasons behind the delays.
Equipped with 16 memory modules offering a maximum bandwidth of 819.2 GB/s, the system is expected to accommodate 1TB of DDR5-6400 SDRAM. Additionally, the configuration includes SATA connectors for SSDs and HDDs, three Gigabit x16 5.0 slots, and three M.2-2280 PCIe slots with a computer-controlled interface.
As an ATX solution, the Prodigy ATX Platform assures comprehensive I/O connectivity, encompassing standard ports such as USB, HDMI, and Ethernet, typical for such a workstation. The motherboard may feature the AST2600 baseboard management controller from Aspeed.
For now, Tachyum has disclosed a schematic layout of the Prodigy ATX Platform’s motherboard and an unoccupied black PC chassis branded with the company’s logo.
The primary focus of the Prodigy ATX Platform lies in leveraging pre-trained models for inference tasks, leveraging the processor’s unique architecture for enhanced efficiency. Tachyum asserts that a model with a trillion parameters utilizing FP8 necessitates 2.04 TB of memory. However, by adopting Tachyum’s 4-bit Chi sparse format with 4-bit weights, the memory footprint drastically reduces to 765 GB, enabling the support of larger models within the system’s 1TB memory capacity.
According to Tachyum, a setup comprising a single 96-core Prodigy processor and 1TB of RAM can efficiently perform inference tasks on the 1.7 trillion-parameter ChatGPT4 model, a feat that would require 52 Nvidia H100 GPUs at a significantly higher cost and power consumption.
Dr. Radoslav Danilak, the founder and CEO of Tachyum, emphasized in a press release that the widespread adoption of Generative AI is imminent, surpassing previous expectations. He envisions AI becoming indispensable for websites, chatbots, and other essential functionalities to ensure a satisfying user experience. The robust AI capabilities of Prodigy empower organizations of all sizes to engage in AI initiatives typically dominated by industry giants, offering faster and more cost-effective operation of Large Language Models (LLMs) compared to existing CPU + GPGPU setups.
Tachyum’s ability to revolutionize the AI landscape hinges on the broader adoption of the Prodigy system, provided an adequate chip supply. However, the scarcity of functional silicon raises doubts about the feasibility of this vision. The timeline for mass production of Prodigy processors and their widespread availability remains uncertain.
Concerns regarding the financial viability of the Prodigy ATX Platform for businesses have surfaced. The inclusion of a highly specialized multi-layer motherboard, an advanced voltage regulating module, a 2000W PSU, and a quality chassis with reliable cooling systems could collectively amount to around \(500. Additionally, the cost of 16 64GB RDIMMs at \)240 per unit totals \(3840. Even excluding the Prodigy processor, the total system cost approximates \)4800.
The economic feasibility of the inference software tailored for Prodigy remains speculative, considering potential cost reductions through bulk purchases. The affordability of Prodigy silicon also plays a crucial role in this equation. However, persistent delays in the product launch raise concerns about Tachyum’s operational timeline.
