According to the company the architecture is intended to simulate the parallel transformation and evolution of quantum-style computation. The company has discussed the architecture at a high level in a press release but not in detail. It does state the SoC architecture includes a RISC-V processor, a QuantaFlow generator and a QF evolution space. QuantaFlow simulates transformation and evolution of registers. The single-core RISC-V processor is used for control and input and output of results. The company said, adding that the transformation/evolution process is highly parallel and asynchronous but provides little other detail.
The company also states that QuantaFlow can run all kinds of neural network including: ResNet-50 (2015), MobileNet (2017), EfficientNet (2019), without speed degradation or hitting the "memory wall."
The performance claim for the QuantaFlow architecture is that it can achieve 10X speedup compared to Nvidia's V100 on ResNet-50 with a single batch and using 8bit integers. For newer network models, performance enhancement should be even greater, the company said.
It is not clear where PQ Labs Inc. is with regard to implementing its architecture in silicon, or where it will get the chip manufactured, on the assumption the company is a fabless chip company.
PQ Labs was founded in 2008 and is provider of multi-touch interface systems using fingers and gestures. The company's customers include the U.S. Air Force, U.S. Army, U.S. Navy, U.S. Department of Homeland Security, Microsoft, AT&T, Xerox, Nvidia, Samsung, LG, Ericsson, Lockheed Martin and others.
Related links and articles: