A company led by a former NASA boss wants to take tricky and compute-intensive algorithms off general-purpose silicon and has just popped out of stealth mode to show off its first efforts.
The de-cloaking, which came after an uncharacteristically long period of ten years in “stealth mode”, revealed to the world a digital signal processor (DSP) designed for signal processing, deep learning and analytics.
Why would it stay beneath the radar so long? Because the company, Knupath, founded by former NASA administrator Dan Goldin, was working in a new architecture.
As discussed in depth at The Register's HPC sister publication The Next Platform, Knupath is betting its future – and US$100 million of patient investment – on “sparse matrix-based computing on low-power and reprogrammable devices”.
Put simply, a sparse matrix is one in which most of the entries are zero, and one of the most famous applications of the sparse matrix is in Google's PageRank.
What Knupath wants is to create sparse matrix-specific hardware that's scalable to more than half a million chips in a system, with 256 “tiny DSP” (tDSP) cores on each chip overseen by an ARM management core. The claimed latency comes from an interconnect produced by Knupath, and the outfit claims an Ethernet-equivalent rack-to-rack latency of 400 ns.
Other headline specs include a distributed memory model shipping 320 gigabytes per second outbound from each chip, and an aggregate 3.7 TB/second whole-of-machine memory bandwidth.
All of this, as Goldin told The Next Platform, rests on the world adopting sparse-matrix approaches to deep learning, and developing architectures specific to that class of problem.
As well as its Hermosa processors and LambdaFabric, Knupath has a developer board program here.
More by Nicole Hemsoth at The Next Platform here. ®