Tilera’s Server Chip Challenges Intel, Sort Of
It’s been awhile since there was a new chip on the scene to get excited about; one that didn’t come from Intel, and wasn’t aimed at a mobile phone. It’s been even longer since there was a chip aimed at servers. Today is one of those days.
A start-up called Tilera today unveiled a chip it calls the TILE-Gx. Essentially, it’s a super-chip with 36 cores which — so the company claims — beats a traditional Intel server chip on the key metric of performance per watt.
For those who don’t keep score in the arcane world of semiconductors, I’ll revisit some of the basics of the above paragraph. We all know that Intel and its one main rival, Advanced Micro Devices, sell chips for servers. Those chips, and those that go into PCs, are generally known as x86 chips, a name derived from the instruction set they share.
On the other hand, there are ARM chips, which are a different breed, and exist in a very different ecosystem. Scores of companies make ARM-based chips for all kinds of different uses, and they license the basics of the designs from ARM, the company, which last year did $636 million in revenue.
ARM chips show up in phones and tablets from the likes of Broadcom, Qualcomm, Texas Instruments and Nvidia, but not so much in PCs and servers. ARM is even the basis for Apple’s A4 and A5 chips. At CES last year, Microsoft said it would create a version of Windows 8 that will support ARM chips. And a company called Calxeda (which I initially got mixed up with Tilera) is aiming to bring ARM cores to chips running in servers.
Tilera, based in San Jose, Calif., is backed by investments from Bessemer Venture Partners, Walden International, Columbia Capital and VentureTech Alliance; plus a trio of strategic investors, Quanta Computer, NTT Finance and Broadcom. Its new chip is based around an entirely new architecture developed by Tilera’s CTO Anant Agarwal, a professor at the Massachusetts Institute of Technology. It forgoes both the traditional x86 and ARM architectures. Aimed squarely at servers, its intention is to get the same work done that a traditional Intel server chip does, while using less power to do it.
That’s not a trivial benefit, especially in data center environments where servers are bunched together and pushed to the performance limit. The biggest operational expense in running them is going to be power. So it’s on this point that server vendors and chip vendors obsess over saving a watt here and there — over the machine’s useful lifetime, the costs will add up considerably.
How it does this is what makes it interesting. Essentially, the cores on the chip do something that an Intel chip can’t do: They communicate among themselves. The way I understand it — and I admit I’m simplifying it greatly — the cores on an x86 chip rely on a single communications channel, called the Bus, to communicate. The Tilera architecture allows each core to communicate directly with the other cores, thus eliminating the need for the Bus and cutting back on the need for power.
The top-end chip — there are two versions — has 36 cores. A core is essentially the main computing engine on a chip. If you’re reading this on a PC, chances are the chip inside it has two cores, maybe four. It used to be that chips had only one core, until it became logical to put two or more on a single chip. I’ve always compared multicore chips to roommates folding laundry together. When there’s a big pile of laundry to be folded, one person can certainly do it, but two or four get it done faster and with less effort. Multicore chips basically prove the old adage that many hands make for fast work.
There’s an obvious appeal to a chip like this, but there are a lot of strikes against it. First, much of the server ecosystem is pretty well entrenched. Companies run what applications they already have, and are usually loath to mess with their computing environments much. Changing the architecture of the CPU chip inside the servers is about as major a decision as a CIO may ever make, and one they don’t make lightly. First they’ll have to test it and run it for awhile, and then see how it interacts with other systems. It’s not the sort of decision that happens just overnight. Also, a new architecture brings with it a lot of software compatibility questions that will give many IT departments pause.
Meanwhile, Intel, which sells chips that go into most of the world’s mainstream servers, will continue to push its power consumption down. At the same time, it’s been trying like crazy to use its Atom line of chips to mount an attack on ARM’s territory and win business from phone and tablet vendors. That effort is just now seeing its first early successes. If there’s a great long-term story in chips that bears watching, the grappling between Intel and the ever-expanding universe of ARM vendors is certainly it.
Correction: I initially thought the Tilera chip was based on the ARM architecture. I’ve revised the story to correct that.