Venture capitalists these days tend to avoid chip companies, largely because developing new products costs a bundle. That hasn’t stopped Soft Machines. The Silicon Valley startup is trying to crack into a particularly tough niche in semiconductors.

The company has been working secretly for seven years on an unusual technology designed to deliver big performance benefits while running software designed for a variety of established types of microprocessor.

It has been an expensive effort. Soft Machines, which disclosed its plans for the first time Thursday, has raised a hefty $125 million in funding and employs 250 people in California, India and Russia. It topped a recent ranking of most heavily funded startups operating in stealth mode when its funding was listed at only $96 million.

Soft Machines plans to license its inventions to other companies designing chips rather than sell its own. It expects initially to target customers exploiting the Android operating system, the mainstay of mobile devices that usually runs on chip designs from ARM Holdings ARM.LN -1.22%. But the company says its technology could also run software written for chips from Intel INTC -0.63%, IBM IBM +0.40%and other microprocessor designers.

“We don’t want to limit ourselves or marry ourselves to one ecosystem,” said Mahesh Lingareddy, the company’s chief executive. “We want to be able to sell this to everyone and we want to be friends to everyone.”

Lingareddy and co-founder Mohammad Abdallah previously worked together at Intel, helping to design chips in an era when designers could exploit steady speed increases that came along with advances in manufacturing technology. Chip operating frequencies, or clock speeds–a measure of the internal pulses that govern computing operations–kept getting faster in the 1990s and early 2000s as Intel production recipes created smaller circuitry.

But Intel and its rivals later changed course, as rising clock speeds created chips that used too much power and generated too much heat. They started exploiting improvements in miniaturization to incorporate more calculating engines, or processor cores, on each piece of silicon.

But few application programs exploit the full benefit of such multi-core chips, Lingareddy says. So the pattern of speed jumps dramatic enough that most users would notice has faded.

Lingareddy, Abdallah and their team tried to respond with an approach that uses the same underlying concept–breaking up computing tasks into small chunks that can be executed simultaneously. But they tried to do it in a different way.

Today’s chips rely on programmers to design products that send separate streams of instructions, called threads, that can be handled by separate computing cores on a single chip. Soft Machines developed a special block of circuitry that automatically breaks up those threads into smaller chunks, working much more quickly, Lingareddy says.

Those instructions are then passed along to what the company calls virtual cores, which are processing engines that Lingareddy says can be assigned specific chores more flexibly than the cores in conventional chips.

There are other nuances and details. But what most people will care about is the results: Soft Machines, which has working samples of its chips, says its tests show that its design can generate two to four times higher computing performance per Watt of power consumed.

That means that a chip designer might choose to run at a standard clock speed and get noticeably better performance, Lingareddy says, or maintain an earlier performance level while sharply reducing power consumption to save battery life.

Industry skepticism is likely. The notion of abstracting software from chip hardware has been tried by companies such as Transmeta, a startup born in the mid-1990s that labored for years in secret on technology based on translating computing instructions in novel ways. The startup ultimately failed.

But Soft Machines seems to have convinced a long list of investors. They include two former senior Intel executives (Albert Yu and Richard Wirt), well-known chip entrepreneur and investor Gordon Campbell, the venture arm of Samsung Electronics 005930.SE -1.53%, Intel rival Advanced Micro Devices AMD -0.56%and Mubadala, the Abu Dhabi investors backing chip manufacturer Globalfoundries.

Lingareddy says he didn’t expect to lean on their assistance quite so much. “If I had known I had to raise $125 million and I had to build a 250-person team, I would have thought twice,” he says.

The company disclosed its plans at an event hosted by Linley Group, a research firm tracking the chip industry.

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved