Arthur Cole speaks with Ihab Bishara, director of marketing, Tilera Corp.

Cole: The new Tile-Gx marks a dramatic increase in the number of cores compared to anything else commercially available. What can you tell us about the architecture that allowed you to pack in so many cores?

Bishara: It's actually very simple. We were able to do it by using standard cores and connecting them with a multi-layered network. We use a 2D mesh network that connects all cores to each other and to a router. We've essentially created a huge internet -on-a-chip that can scale to hundreds of cores, not just 100.

In designing the architecture, we focused on making sure each core has enough bandwidth, so we devoted more than a TB per core. This is so you can have multiple messages flowing around in the chip without blocking each other. This is a different approach from current multicore processors, where they take a standard single core, put a bunch of them together and connect them to a bus. The core they are using doesn't lend itself to multicore operation because it is the legacy core we've been working with for years. And the bus is a 1D system, so if one core wants to control it, everything else must get off, which bottlenecks the whole system.

We also have a distributed cache system, where each core can access its own local cache or any other cache on the chip.

Cole: You've mentioned cloud environments as a key application. What sort of functionality could it bring to server, storage and networking components?

Bishara: Power and cooling are two of the toughest problems that data centers face today. We have data centers running of out power even though they have lots of empty space, and data centers being put next to dams for cheaper and more power, or underground to make them cooler.

Today, with racks and racks of servers, you can take five or seven boxes and condense them down to one box with the same performance. (Current processing technology) has a huge advantage in the data center market, and it has become so entrenched that it will be hard to unseat. That's why we don't focus on general-purpose enterprise servers.

But we do focus on the cloud data center where much of the architecture is very simple: You have some memory cache with various databases like MySQL, plus some number-crunching applications like Hadoop. So you have dozens of applications running over a million servers. That's where our focus is: creating specialized servers and optimizing applications on a performance-per-watt level.

Cole: Much of the software industry is still struggling with multi-threaded programming. How does your Multicore Development Environment help them cope with the added threading capability of the Gx?

Bishara: Multicore programming is hard when you start with huge sequential apps like Windows. But if you look at the applications that target the embedded or cloud markets, by nature they are extremely parallel -- dealing with thousands of users or video streams at once. Most of our customers take an application, put it through a load balancer to balance it across all cores, and each core executes the entire application. By default there is independence among flows. You don't have to take an application and split it across cores, each one processing its own little piece.

We made sure to use standard C++, with a standard compiler, which is what most of our users grew up with. We've enhanced that to make it more multicore-friendly, particularly in debugging. We offer a complete view of all cores and the relationships between them, and we provide a substantial amount of profiling to see where more or less parallelism is warranted.