Tachyum announced its Prodigy ATX platform on Friday in a special white paper. It’s an artificial intelligence-oriented workstation based on a cut-down version of his Prodigy universal processor that still needs tape-out. The company plans to sell the unit for $5,000 and claims the machine will democratize access to large language models with billions (or trillions) of parameters. However, the company has not disclosed when the product will be released.
The Prodigy ATX platform is based on the 96-core Prodigy processor (scheduled for mass production on TSMC’s 5nm node in 2024) operating at up to 5.70 GHz. This CPU is said to have only half a die to reduce power consumption and improve yield, which will reduce cost and make the platform more accessible, Tachyum said. Stated.
As for this processor, Prodigy’s release was initially set for 2020 after a tapeout in 2019, but it has been pushed back several times with increasingly extravagant performance claims without a proven prototype. It is worth mentioning that it has also been postponed to 2021, 2022, and 2023. Currently, the company’s latest plans call for the launch of Prodigy processors in late 2024, which could mean December 2024, although no details have been disclosed.
The machine will be equipped with 1TB DDR5-6400 SDRAM using 16 memory modules and offering a peak bandwidth of 819.2 GB/s. The system will have three PCIe x16 5.0 slots, three M.2-2280 NVMe slots with PCIe 5.0 x4 interfaces, and SATA connectors for SSDs and HDDs.
Being an ATX box, the system promises to offer all the I/O connectors you’d expect from a machine like this, including USB, HDMI, and Ethernet. Additionally, the motherboard will be equipped with Aspeed’s AST2600 board management controller.
For now, Tachyum has released a block diagram of the Prodigy ATX Platform motherboard and an empty gray PC chassis with its name on it.
The focus of the Prodigy ATX platform is to use pre-trained models for inference and make inference more efficient due to the architectural characteristics of the processor. According to Tachyum, when using FP8, he needs 2.04 TB of memory for a model with 1 trillion parameters. However, by employing his 4-bit TAI sparse with Tachyum’s 4-bit weighted format, his memory requirements have been significantly reduced to his 765 GB, making the system even larger within his 1 TB limit of memory. model can now be supported.
Tachyum says a system with a single 96-core Prodigy processor with 1TB of RAM can run inference on a ChatGPT4 model with 1.7 trillion parameters, but it would take 52 to do the same thing. I mentioned that it requires an Nvidia H100 GPU, which is significantly more costly and power-consuming. . ”
“Generative AI will become widely used much faster than anyone initially expected,” Dr. Radoslav Danilak, founder and CEO of Tachyum, said in a press release. . “Within 1-2 years, AI will become a required component of websites, chatbots, and other critical productivity components to ensure a great user experience. Prodigy’s powerful AI capabilities will enable , LLM is much easier and more cost-effective to run than existing CPU + GPGPU-based systems, allowing organizations of all sizes to compete on AI initiatives that would otherwise be dominated by industry giants. It will end up happening.”
If Tachyum succeeds in popularizing its Prodigy platform, it could actually change the rules of the game in terms of AI, provided it can supply enough processors. However, it is unclear whether this can actually happen, given the lack of functional silicon. Furthermore, it is unclear when exactly Tachyum will begin mass production of its Prodigy processors, or when it plans to make them available in large quantities.
There are some reasonable doubts about the company’s financial viability as far as the Prodigy ATX platform is concerned. Sixteen 64GB RDIMMs at $240 per unit cost $3,840, and a highly customized multi-layer motherboard (we’re talking 12-16 layers here) with advanced voltage regulation modules ends up costing $500. (we’re talking about relatively small amounts of product). A batch of motherboards, these are expensive), a 2000W PSU costs around $300, and a decent chassis with a reliable cooler costs around $150. Even without Tachyum’s Prodigy processor, the system costs around $4,800 in total.
Of course, the company buys things in bulk, so it ends up costing much less, but Prodigy’s silicon is probably also expensive, making a platform designed for inference economically viable for Tachyum. I doubt whether it will be a good product. And of course, there’s also growing skepticism over the company’s often delayed launch schedule.