SANTA CLARA, Calif. — Two former Intel engineers who faced the power wall that ended frequency scaling are describing their startup’s technology that hopes to surmount it. Soft Machines aims to bring a new level of parallelism to single-threaded software that could enable performance gains for any processor.
The startup has attracted $125 million in financing since it was formed in 2008, making it one of the few well-funded processor startups these days -- and certainly not the fastest to get to market. It hopes to close soon a “huge” round of investment that will take it through the next three years when it aims to help its first customers bring working chips to market.
The company is showing a working prototype of a 32-bit processor that uses its microarchitecture and software to provide about twice the performance of 32-bit ARM and x86 chips. It delivers about 40% more than Intel’s current 64-bit Haswell processor, the startup claims.
Overall, the company claims its approach can deliver the same performance at a third or a quarter of the power of conventional processors or 1.7x to 2.2x the performance at the same power level. It claims the approach can be used on applications ranging from embedded devices to datacenter servers.
Soft Machines is working on a handful of implementations for different markets of its so-called Virtual Instruction Set Computing (VISC) technology. It aims to forge deals with customers next year to license the technology and co-develop processors using it.
Chief executive Mahesh Lingareddy and CTO Mohammad Abdallah designed processors together at Intel for eight years ending in 2006. Their experiences led them to form Soft Machines, which got angel funding from, among others, two of their former bosses at Intel, Richard Wirt and Albert Yu.
“We saw first-hand the power wall, working on a 6 GHz design in early 2000 that aimed to evolve to 10 GHz in 2010 -- the rest was history,” says Lingareddy. “As the industry moved to multicore, we realized there would not be the same power-per-watt scaling because of programming challenges and other issues.”
Soft Machines claims its approach can split a single-threaded application into separate tasks that run across a pipeline of virtual threads over one or more virtual cores. The virtual cores can be mapped to resources on one or more physical cores.
The Global Front End is a hardware scheduler that sends instructions to different virtual cores. The underlying physical cores have some unique communications and synchronization hardware to facilitate the work.
The approach is more fine-grained than today’s multi-threading technologies, says Abdallah. It’s also more efficient than virtual machines because it is based in hardware rather than high-level software constructs.
There is some latency and overhead in the approach, but the net returns in simplifying the complexity of an out-of-order pipeline make it worthwhile, he claims.
The startup described its technology at the Linley Tech Processor Conference here. It aims to announce products and first users sometime next year. Its backers include AMD, GlobalFoundries, and Samsung, as well as national investment groups in the Middle East and Russia.
Next page: Startup shows benchmarks on its prototype
— Rick Merritt, Silicon Valley Bureau Chief, EE Times