Tachyum, which has been promising a "universal processor" dubbed Prodigy for the past three years, said it was one step closer to delivering on that pledge with the development of a motherboard for its FPGA emulator that allows customers to test a complete Prodigy system.
Formed in 2017 by Skyera and SandForce co-founder Dr Radoslav "Rado" Danilak, Wave Computing co-founder Ken Wagner, engineer Igor Shevlyakov, and hardware architect Rod Mullendore, Tachyum was straight out of the gate with a bold projection for its hardware: a tenfold performance boost.
Tachyum was funded on the promise of what it described as the "Cloud Chip", which was formally announced in 2018 as the Prodigy – an alternative to Intel's server-centric Xeon family of processors with its own instruction set architecture and compatibility with ARM and x86 legacy software provided through an emulation layer, yet somehow running faster than when executed natively on a Xeon chip.
- Prodigy chip brainiacs Tachyum hook up with Euro HPC consortium
- Xeon-bashing Tachyum claims its Prodigy CPU will run AI jobs as well as traditional apps
- 'Prodigy' chip moonshot gets hand from Arm CPU guru Prof Steve Furber
- Ex-Skyera chaps' Tachyum startup grabs cloud chip cash
At the time, the novel chip design was said to be heading to tape-out and manufacture at TSMC in 2019. It's now 2021, and there's no Prodigy. Despite joining an EU project to build an exascale supercomputer, Tachyum has not yet built a commercially available Prodigy in silicon – though it has, it says, proven the concept on a field-programmable gate array (FPGA).
Now that same FPGA prototype can be installed into a functional computing system. "Our IO motherboard, in conjunction with our CPU motherboard, enables our engineers to fully test the functionality of Prodigy," Danilak announced of his company's latest hardware.
"Together, these two FPGA-based boards provide the basis of a system that can be cascaded to fully emulate an entire 128-core Prodigy processor, which is capable of advancing the entire world to a greener era by enabling human brain-scale AI."
The FPGA prototype has been made available to "early adopter partners," the company confirmed, with the project on track for sampling the whole Prodigy system later this year. Feedback from testing with the FPGA prototypes will feed back into tape-out on an aggressive 5nm process node, Tachyum has claimed. A four-socket motherboard, meanwhile, is scheduled for production by year's end.
Despite still working with power-hungry FPGA emulation, Tachyum is sticking to its headline claims: a tenfold reduction in power draw core-versus-core, the ability to beat Nvidia's "fastest GPU" in high-performance compute workloads, AI training, and inference, and a fourfold decrease in total cost of ownership compared with established rivals.
Whether the Prodigy will deliver on these claims, and the promise of the ability to switch between workloads written for dramatically different architectures on the fly while running them faster than native, has yet to be independently verified. Interested parties, however, can apply for access to the emulation system on the company website. ®