Tachyum Claims Its 1024 Core, 6GHz Prodigy AI Chip Crushes NVIDIA's Rubin Ultra

Tachyum is back with another round of astonishing promises. The startup, which was founded in 2016 and still yet to ship a single chip, has unveiled the latest revision of its Prodigy processor design, now targeted at 2nm, which it says delivers up to 21× higher AI rack performance than NVIDIA's future Rubin Ultra platform. The catch? Neither Prodigy nor Rubin Ultra actually exists yet. This is, in effect, children on a playground arguing whether Goku could beat Superman, only the playground costs billions of dollars in wafer starts.

According to Tachyum's press release, the top-end Prodigy Ultimate model packs 1,024 custom 64-bit cores running at up to 6GHz, supported by 24 DDR5-17600 memory channels and 128 PCIe 7.0 lanes. The chip is said to exceed 1,000 PFLOPs of inference performance per rack and run circles around NVIDIA's upcoming data-center GPUs, all while drawing a mere 1,600W per socket.


To be frank, if a startup that's never shipped silicon could pull this off, it would make AMD, NVIDIA, and Intel look like they've been asleep at the wheel. The company says it has "secured tape-out funding," which is another way of saying the design hasn't been fabricated yet. Every quoted performance number comes from internal simulations, not real hardware. That's totally normal; NVIDIA's quoted performance numbers for its Rubin hardware are largely based on simulations too. The difference is that we believe NVIDIA can actually ship those specs based on its history.

Meanwhile, even the specs themselves for Prodigy raise eyebrows. DDR5-17600 simply doesn't exist, neither in JEDEC specifications nor in bleeding-edge overclocked modules. Even if it were to use soldered-down LPDDR5X, the fastest LPDDR5X from Samsung peaks at 10.7Gbps per pin. Likewise, PCIe 7.0 is still in the verification phase, and won't see real deployments for a few years yet. In other words, Prodigy's feature sheet reads more like a wish list than a bill of materials.


There's a reason the world's fastest supercomputers, the DOE's Aurora, Frontier, and El Capitan machines, all pair CPUs with massive GPU accelerators instead of relying solely on CPUs. Even NVIDIA's NVL racks use a mix of Grace CPUs and Blackwell GPUs. You need CPUs to prepare the work and manage coordination between nodes, while special-purpose matrix hardware delivers far better throughput and efficiency for modern AI workloads. Tachyum's dream of a "universal processor" that replaces CPUs, GPUs, and TPUs with a single design is elegant in theory, but it's one the rest of the industry has already tested, costed, and moved on from.

That said, Tachyum's ambition isn't completely without merit. The idea of unifying compute architectures is appealing, and the team has legitimate engineering credentials. Beyond founder Dr. Radoslav Danilak, whose SSD-controller and GPU/CPU architecture talent are well-documented, the company also lists Pini Herman, formerly of Cadence and HGST, as Senior Director of Solutions Engineering, underscoring that the team isn’t entirely green. Still, until a foundry turns the design into working silicon—and someone outside Tachyum's marketing department measures it—these "21×" claims belong in the realm of simulation, not science.

For now, Prodigy remains a bold PowerPoint project promising to outrun NVIDIA's next-generation chips. If it ever materializes, we'll all be delighted to eat our skepticism. Until then, this is still Goku versus Superman.