Chip Startup Tilera Dreams the Impossible Dream


TILEPro_processor_pageAnant Agarwal, co-founder and CTO of Tilera, is tackling the Mount Everest of chips. His goal for decades has been to figure out how to build a general-purpose chip that offers better performance and power efficiency. Those are goals that many startups have shot for and missed, or aimed for and settled for two out of three. But like Everest’s awesome heights, the new generation of cloud computing and demand for ever more resources to power social networks, online video and devices, have created challenges for data center operators that may allow Tilera to succeed. Today, the 5-year-old startup is expected to launch a 100-core version of its chip aimed at web-scale computing.

Tilera scoffs at quad core machines. The company’s chips already are used by 75 customers, and come with 36, 64, and now 100 cores. Agarwal says, “The core is the new transistor.” By cramming so many cores onto its chips connected by a mesh network of interconnect that allows the cores to communicate without bottlenecks, Bob Doud, director of marketing, says that Tilera can sell its chips to folks wanting faster memcached servers or better performance at web-scale computing. The chips, which provide 1.25 GHz of performance, are no match for Intel’s (s intc) workhorse Nehalem processor that can top out at 3.3GHz. But Tilera’s chips only burn 33-50 watts instead of 130 watts that top-of-the-line Nehalem silicon can.

Giving folks better performance per watt has an essential place in the world of web-scale computing (GigaOM Pro subscription required), as Microsoft (s msft) tests Intel’s low-end Atom chips in its servers; Dell (s dell) touts servers using the low-power VIA chip; and SeaMicro creates an 80-core Atom-based box. Doud sees this opportunity and hopes that Tilera’s unique architecture can outperform efforts by other startups (and even big chip vendors) and get inside boxes made by large OEMs such as Dell and HP (s hpq) or smaller vendors using x86 chips in their cloud appliances.

But it’s this architecture issue that’s Tilera’s biggest weak point. The difficulty for most startups in the chip space when trying to build computers that deliver the most performance for the least power is the fact that they’re competing against Intel, and huge chunks of code aimed at enterprise and personal computing are written for Intel’s x86 architecture. So even as established chip providers (such as Nivida (s nvda), which has had some luck getting its graphics processors into the mainstream) try to get people to write programs for their GPUs, Tilera has to offer tools to help programmers write for its chips without learning a new programming language.

Doud says he thinks 10 percent of web-scale computing jobs don’t need to keep the old style of coding for x86 chips, and if Tilera can break in there, that would be enough for now. Tilera already sells its chips to telecommunications equipment companies and for those trying to do rapid video and voice transcoding, which Doud says helps the company diversify in a manner that other specialty chip vendors, such as the shuttered SiCortex,  have not done.

As Agarwal takes Tilera up the chip industry’s version of Everest, he is betting that the huge shift that’s come about as a result of web-scale and cloud computing is a good time to challenge the need to keep writing for x86 chips. The startup has made it to base camp, with an expected $25 million funding round from computer manufacturer Quanta and undisclosed investors closing later this month (bringing its total funding to about $65 million), and a plan to become profitable without any new investment, but it’s still going to be a hard, uphill slog.


Brett Glass

The author of this article, like most of the public, apparently doesn’t understand the basic technical fact that “gigahertz” is not a measure of computing power.


I’d say a little differently…

Its not THE measure, but it is A measure. Where this article should be more focused is the apps and perf metrics related to those apps. Benchmarks relevant to the PC app universe and telco/web universe are different and it would be useful if Tilera (more than the author) can tout some of those..

Stacey Higginbotham

I am aware of the limitations of GHz as a metric, but it’s also an easier and more familiar way of making an apples to apples comparison for readers. The bottom line is that as price performance per watt becomes more important, and if hardware specific clouds evolve, then we can measure performance on a per-job basis and apply the perfect hardware for the job. I’m not 100 percent sure this will actually happen, but the energy and cloud trends make this seem plausible.

Comments are closed.