The power of 52 NVIDIA H200 accelerators at a hundredth

The power of 52 NVIDIA H200 accelerators at a hundredth of the cost – Tachyum will begin mass production of 192-core Prodigy chips this year

Tachyum announced that this year it will begin mass production of the general-purpose processor Prodigy, which combines computing and graphics cores and a TPU neural engine. The manufacturer promises that the new product can offer an outstanding level of performance and compete with the most powerful NVIDIA computer accelerators.

    Image source: Tachyum

Image source: Tachyum

The developer claims that its 192-core chip, based on 5nm process technology, delivers 4.5x faster performance than any other processor designed for cloud computing workloads. Additionally, it is up to three times faster than any GPU designed for high-performance computing and up to six times more power efficient than dedicated AI GPUs.

Tachyum announced the Prodigy Universal Processor Chip in 2022, promising to transform highly specialized data centers into general purpose data centers that can provide the computing power and efficiency necessary for a variety of AI workloads. In December 2023, the company released a video showing Prodigy’s ability to emulate x86-compatible applications. However, there is one major “but”: despite the impressive performance numbers and Tachyum’s claims, Prodigy processors only exist on paper and as an FPGA-emulated platform with a small number of cores.

“The progress we have made in adapting our product roadmap has us excited as we move into 2024 as we move toward mass production of Prodigy and a multi-billion dollar distribution channel. We look forward to delivering on our promise and commitment to transform traditional data centers into all-in-one data centers in the near future.”said Radoslav Danilak, CEO and founder of Tachyum.

Because the Prodigy Universal Processor uses functional components designed for different types of workloads, it can dynamically switch between computing clusters, eliminating the need to use different and expensive hardware for specific types of AI workloads within a computer system. At least that’s what the company’s latest press release says.

It also says that only a Prodigy general-purpose processor, costing $23,000, can match the performance in AI training tasks with a system of 52 specialized NVIDIA H200 accelerators, which are among the best on the market. The company said the cost of that many GPUs across seven Supermicro GPU servers would be $2,349,028, 100 times more than a system with a Prodigy general-purpose processor and 2TB of DDR5 RAM.

Tachyum also stated that it is already planning to develop a more advanced general purpose processor Prodigy 2 based on 3nm process technology, which supports PCIe 6.0 and CXL interfaces and will be equipped with integrated high-performance HBM3 memory. Publication is scheduled for 2026.

About the author

Dylan Harris

Dylan Harris is fascinated by tests and reviews of computer hardware.

Add Comment

Click here to post a comment