This article is more than 1 year old
VISC-y business: Can Soft Machines keep the free lunch counter open?
It's a multicore world alright, but the ghost of Transmeta lingers
If the free lunch is over, then few have heeded the end-of-break bell.
A decade ago next year, C++ expert Herb Sutter wrote that those building the world’s software would need to change they way they architect and construct.
Multi-core and hyperthreading were incoming, meaning that apps could no longer be built using sequential commands intended for single cores.
Harnessing the performance gain of the new chips meant going concurrent or parallel with instructions run sequentially, in different parts of a chip.
The joke in computing used to be that what Andy giveth, Bill taketh away – each time Intel (under then CEO Andy Grove) brought out a faster chip, Microsoft would embiggen its software to gobble up the new chip’s power.
10 years on, multi-core is ubiquitous, having migrated from the limited uplands of PowerPC and Sparc IV servers to the vast plains of smart phones.
But still, most or today’s software continues to be built to execute sequentially not on parallel.
The problem with parallel is it’s so damn complicated.
Most parallel programming is done using “explicit parallelism” - where you know the architecture and your application is built specifically for that set.
It’s not a scalable process, meaning if the underlying hardware changes or if you want to target more than one maker’s chips, you need to re-write.
Hoping to bring parallel programming to the masses is Soft Machines, a start up that slipped out of stealth mode in October and in 2015 plans its first product.
Soft Machines reckons it has devised an architecture – VISC – that takes care of the parallelism for you – a type of implicit parallelism.
Soft Machines is the brainchild of two former Intel employees, Mahesh Lingareddy and Mohammad Abdallah.
Their VISC is hardware and software that interprets commands coming in and looks for ways to distribute them as “threadlets” – or virtual threads – over a chip’s cores.
Their goal is to license VISC to chip makers who will bake the technology into their own boards and silicon.
Soft Machines told The Reg it will talk licensees, roadmap and products in 2015. At the moment, VISC is a reference board and a prototype SoC, we were told.
VISC, the firm said, will know the ins and outs of the host chip’s multi-core architecture and be aware of things such as dependencies and availability of cache – things the incoming software won't know about.
VISC uses Soft Machines' instructions to make devs' code run in parallel with no need for programmers to learn their code, the company said. Rather, a software layer converts binaries such as those for ARM or for x86 into binaries using Soft Machine’s own, internal instructions.
The software also creates a virtual core, containing all the states associated with each thread – elements such as execution and resources.
The idea of hardware acceleration and intelligence in the software isn’t new to Silicon Valley’s chipheads, the question is how far it can be a broad success.
Transmeta was the rock star of the 2000s: the fabled semiconductor company who built x86-compatible processors using VLIW core and what it called Code Morphing Software, which would optimize performance and execution.
Transmetta made two processors, Crusoe and Efficeon but the business faltered and eventually Transmetta went into IP licensing before being acquired in 2009: Novafora bought Transmeta and Intellectual Ventures the patents portfolio.
The dominant chipmakers neither feared nor needed Transmeta, so the firm struggled and was finally crushed.
Soft Machines got going in 2006, during Transmeta’s heyday and just after Sutter’s free-lunch call but it was only in 2010 when things got really serious.
It wasn’t until then that Lingareddy and Abdallah got the funding that let them go beyond just 20 employees and surviving on their credit cards.
Today they have 250 staff and $125m from seven investors, including AMD and Samsung Ventures. The firm has 80 patents to its name.
Abdallah, also president and chief technology officer, told us he and Lingareddy bet their careers on Soft Machines and VISC and it’s taken time for the industry to catch up. He believes there's still a bright future ahead.
“At Intel, even when they had the Pentium 4 or Itanium, it took eight to 10 years and thousands of people to bring something out and they spent more than $1bn,” Abdallah said.
How does this nearly 10-year-old startup succeed in hitting the mainstream and avoid the fate of Transmeta?
Licensing VSIC is the key - thereby working with the system rather than trying to compete against it, according to Abdallah.
Soft Machines wants to license its kit to ARM licensees and those in the Intel and AMD space on x86.
“Transmeta tried to take the market on their own product. For us, we are licensing a technology that can live and work with other architecture ecosystems,” he said.
The other thing is the architecture that, according to Abdallah, lets you run your existing sequential apps without changing for multi-cores and threads.
“VISC lets you scale the number of cores and still run a single multi- threaded application. You can scale the number of physical cores and virtual cores to go the same design,” he said.
Also, VISC delivers linear rather than quadratic scaling between power and performance, meaning high performance for power.
That could pique the interest of those in mobile and devices, where you need more bang for a low-power buck on a growing number of cores.
If Abdallah is right, you might want to hold off ringing that lunch bell a little longer. ®