A new well-funded startup called "Soft Machines," created by chip veterans, is working on a new type of processor based on the "Virtual Instruction Set Computing" (or VISC, as opposed to RISC or CISC) architecture. The firm is essentially creating "virtual cores" that it says are 1.7-2.2 times faster for a given single-threaded task (scoring 2.1 instructions per core in SPEC 2006 benchmark) compared to Intel Haswell's 1.39 instructions per clock.
Most apps or programs aren't usually optimized for multi-core processors or multiple threads, so they end up using a single core, even if the processor has more. That's wasted performance, and wasted time, as the processing could happen much faster if it used all or most of the cores.
The way VISC works is that it takes multiple physical cores to create a single "virtual core," which now has higher single-threaded performance than a single physical core. The main benefit here is that the apps would only use a single thread in the first place. These apps can continue to use one thread, but performance-wise they will perform as if they were using multiple threads or cores for processing, or a significantly faster physical core.
Soft Machines claimed that if its virtual cores maintained comparable performance to regular processor cores, its chips' power consumption could be reduced by one quarter to one third. The power consumption could drop by half if the app would be using two virtual threads. The virtual chip can apparently split into multiple virtual cores, as well.
The technology prototype is already able to boot Linux and Android 4.0 (Ice Cream Sandwich). The company said it eventually wants to build a complete System on a Chip with a GPU, video accelerator and DRAM controller.
What Soft Machines is doing sounds a little similar to what Nvidia did with its Denver core. Nvidia scrapped hardware-based out-of-order execution in favor of an in-order hardware design with software-based out-of-order-execution.
Thanks to this new design, Denver achieves the highest single-threaded performance in the mobile chip market right now, going by Geekbench benchmarks. It's also important to note that Geekbench only supports 32-bit Android right now (so a further 10-20 percent performance increase is expected for the 64-bit mode), and that Denver is still a 28nm chip, while some of its competition (Apple, Samsung) has already moved to 20nm. Denver on the same 20nm could have seen even higher performance.
It's usually close to impossible for a chip start-up to succeed in a world of chip giants such as Intel, AMD, Qualcomm and Nvidia, but if Soft Machines can't make it on its own, it might find a good home at Nvidia, which has just begun experimenting with software-based CPU designs.