Startup processor aims to replace both CPU and GPU with unified chip architecture.
Just over a month ago, Tachyum, a small tech startup, announced a new processor family called Prodigy, which it claims is the world's first universal processor platform. The company has made several striking assertions about the technology, including that it delivers a 10x performance improvement on conventional processors. According to Dr. Radoslav 'Rado' Danilak, Tachyum's CEO, Prodigy represents a fundamentally new approach to processor design. The architecture features a single programming model, a single instruction stream, fully coherent memory, and fully coherent inter-core communication, with added data parallelism to better address certain AI applications.
Prodigy is engineered as a single unified architecture that exhibits out-of-order performance while maintaining processor hardware characteristics—in terms of transistor counts and core size—similar to simple in-order execution machines. Tachyum has accomplished this by offloading to its compiler tasks traditionally implemented in CPU hardware, resulting in improved IPC, clock speed, and power reduction. The processor exceeds NVIDIA Volta performance on neural nets and demonstrates what Tachyum claims is an order of magnitude better performance on Symbolic AI, Bio AI, and General AI compared to existing AI-centric chips, due to their control-intensive nature. The Prodigy architecture includes architectural improvements such as compressed 8-bit floating-point coefficients and matrix multiply-add operations to enhance performance on AI workloads.
Dr. Danilak's success with Prodigy draws on decades of processor design experience, including work on the Playstation 2 and Tesla processors, as well as flash memory controller design at Sandforce and flash-based systems at Skyera. His design philosophy for Prodigy centers on reducing the number of slow wires on a chip and reducing the average length of existing wires—an approach driven by the reality that processor clock speeds have stagnated for a decade due to slow wires relative to transistor switching speed, coupled with CPU architectures designed when wires were considered infinitely fast compared to transistors. According to Danilak, breakthrough performance and low power consumption emerged from this clean-sheet approach.
Danilak attributes Tachyum's potential to succeed despite smaller scale than competitors like Samsung, NVIDIA, and Intel to technical instincts born from hard-won experience, learning from competitors' mistakes, and disciplining focus on important challenges. He notes that at SandForce, the company he founded, his team of less than 100 employees competed successfully against companies like Intel, Samsung, Toshiba, Sandisk, Micron, Western Digital, LSI, and Seagate, each with thousands of engineers. He emphasizes that real innovation in large companies usually springs from small groups of innovators, and points to examples like NVIDIA, where key innovation was done by a couple of groups with ten engineers, and Intel's failed Itanium project, which he suggests failed on compilers—a lesson that led Tachyum to develop compilers first and build the architecture around them.
Tachyum's 64-core Prodigy processor will generate approximately 128TFLOPS of peak performance. The company has suggested that building a candidate processor for the human brain project would require approximately 250,000 of its chips.