Larrabee: What we know so far

Atomic wants its 32 core processor. Here's why, and how Intel plans to take on Nvidia and AMD on the graphics front

Larrabee is a multicore x86 chip that has been optimised for throughput computing.

What this means is that unlike standard x86 chips, the focus isn’t on the time it takes a single program to complete, it’s about how long it takes for lots of invocations of the same program to complete. It's power by a different yardstick.

It's come about because Intel engineers found out that you can get lots of cores on a die the size of a Core2 Duo if you use processors based on a stripped back Pentium architecture. That is assuming you're using modern lithography manufacturing processes. So while initial Larrabee designs will probably feature 24 or 32 cores, as we start to see chips rolling out of the fabrication plants at 32 nm processes and below, we’ll see Larrabee designs with more cores.

We hear you asking what it looks like -- conceptually at least.



All the cores are connected to a shared L2 cache through a wide ring bus, and aware of the entire cache, so any core can access data that isn’t being written to by another core. The L2 cache prefetches data from memory, as does each core’s own 32KB of L1 cache from the L2 cache.

The cores have some changes from the original Pentium design, most notably the addition of a 16-wide vector processing element. Vector processing is the concept of going all-out parallel, like the PS3’s cell processor. You don’t get the raw serial power of the Core2 architecture because the cores are missing several powerful features -- like out of order execution and register renaming. But because of the wide vector processors on Larrabee’s modified Pentium cores, the parallel performance is vastly better than the Core2 design.

And it’s for this reason that Larrabee -- this strangely parallel X86 processor -- is initially going to be used as a graphics processor. The large number of small x86 compatible cores are conceptually similar to the stream processors in graphics cards, but can be programmed with more ease and familiarity and performance should scale according to how many cores Intel manages to cram into the die.



Larrabee is going to take on the high end graphics market but will go about it in a strange way that tips its hat to the old software rendering days of yore. Unlike current graphics cards which pass render frames by passing them through a series of through a pipeline comprised of a combination of fixed function and programmable rendering stages, the stages in graphics rendering pipelines are all emulated.

This means that Larrabee has the potential to be compatible with every DirectX and Open GL version that has been and will be created, because you can set up the chip’s pipeline to emulate the different stages of computation and allocate resources according to the characteristics of the game’s code.

It’s also technically possible to reconfigure Larrabee’s layout as a game is running – for instance, you could optimise the amount of horsepower that goes into building individual frames depending on what the frames require. Beef up the geometry for a series of geometrically complex frames, and then reconfigure for the next frame.

Interestingly, this means that Larrabee even has the potential to bring back the software rendering pipelines by running them in hardware.

We’ll have more for you after we’ve heard more from Intel and have understood what we’ve heard. This is a major and complex shift in computing, and one that has the potential to change the rules of parallel code and data workflows as we know it. We need to know more about what it's like to program for before we come to any decisions about a product that won't even be available for what will probably end up being a year.