COOL-0: Design of an RSFQ Subsystem for Petaflops Computing

We discuss a preliminary design of a Rapid SingleFlux-Quantum (RSFQ) subsystem for general-purpose computers with petaflops-scale performance. The subsystem is being developed at Stony Brook within the framework of the Hybrid Technology MultiThreading (HTMT) project. COOL-0 design is based on 0.8-μm RSFQ technology which enables the implementation of superconductor processing elements (SPELLs) operating at clock frequencies up to 100 GHz, pipelined cryo-memory (CRAM) with 30 ps cycle time and interprocessor network (CNET) with a bandwidth of 30 Gbps per channel. The main architectural challenge is an almost 1,000fold speed difference between the RSFQ processors and roomtemperature SRAM comprising the second level of the HTMT memory hierarchy. The proposed solution to the problem is hardware support for two-level multithreading and block transfer techniques in SPELLs. Our preliminary estimates show that an RSFQ subsystem with 4K SPELLs and a 4-Gbyte CRAM may be sufficient to achieve the performance close to 0.5 petaflops for computationally intensive program kernels. COOL-0 would occupy a physical space of about 0.5 m³ and dissipate power as low as 250 Watts (at helium temperature). These numbers present a dramatic improvement compared to a hypothetical purely-semiconductor petaflops-scale computer.