'The CPU is the weakest link in computing'—Flow says it can boost CPU performance up to 100x with its proprietary companion chip

pcgamer.com:

CPU design has been undergoing something of a shake-up recently, what with the addition of NPUs into the mix and some radical configurations and chip construction techniques to optimise them. Flow Computing, however, has an even more radical suggestion. It's touting a co-processor architecture that gives regular old CPUs a helping hand, with claims of up to a 100x performance increase after some software trickery.

Flow is a startup backed by Finland's state-owned VTT research organisation (via Techcrunch), and its proposal is to add a Parallel Processing Unit (PPU) integrated into a co-processor that works alongside a traditional CPU to help traffic-manage incoming tasks. 

Modern, multicore CPUs can handle dozens of threads at a time, but synchronising them all and handling the data required takes up lots of valuable clock cycles. Flow's PPU design essentially handles this for the CPU, reducing latencies, especially in massively parallel workloads.

Flow's co-founder and CEO Timo Valtonen said: «The CPU is the weakest link in computing...it’s not up to its task, and this will need to change.”

This isn't the first time such an idea has been suggested. Traditionally, however, this sort of solution would require all code to be completely rewritten to facilitate the process. Flow says that its architecture can enhance legacy code without alteration, significantly boosting performance, particularly in regard to apps that display parallelism but are held back by traditional thread-based processing.

That's not all, though. With some software recompilation, Flow says that modified code built to take advantage of the PPU would be capable of performance increases of up to 100x, and has posited the release of custom tools to help developers adapt.

Production model PPUs are proposed to be produced in multiple configurations for different use cases, with differing numbers of PPU cores. A four-core version is said to be suited to something like a smartwatch, while the company