Intel's brand new graphics chip is made of lots of ancient Pentium processors. Is that potty or pretty prescient?

Spare a thought for AMD and NVIDIA. They've been happily smacking each other upside the head for a decade. But at least they've been doing so safe in the knowledge that their GPUs are distinct from - and inherently superior in graphics processing terms - to CPUs.

At least, that used to be the case. Intel has now unveiled Larrabee, a co-processor based on an entirely new approach to graphics processing. If Intel has done its sums correctly, not only will the very definition of the GPU be unceremoniously defenestrated, but also AMD and NVIDIA's graphics chips could even be pummeled to the very brink of existence.

If that sounds like hubris, try this for size. Larrabee could also tear up the rulebook for CPUs, too. That's right, a single new architecture might just take over as ultimate all-round processor, eventually cannibalizing sales of Intel's own conventional CPUs.

What exactly is this deathly destroyer, this harbinger of doom made manifest in 45nm silicon? Some sort of retro-engineered alien technology? Well, here's the really hilarious bit. In simple terms, it's just a metric crapload of old Pentium MMX processors crammed into a single processor die and mounted on a PCI Express board much like any other graphics card.

Exactly how many of these cores Larrabee contains Intel will as yet not be drawn on. Given that chips based on the Larrabee architecture won't go on sale until late this year or early in 2010, it's entirely possible Intel has yet to finalize the core count.

Watching the clock cycles

However, we do know that the Pentium-derived design of the cores makes them much smaller than those found in an Intel Core 2 die. In fact, Intel suggests 10 Larrabee cores can fit in the same space as a single 65nm Core 2 Duo die. Extrapolate out from that using the knowledge that Larrabee will be based on 45nm silicon technology and we reckon Larrabee chips will boast at least 32 cores at launch.

Other than the Pentium link, the other major feature of the Larrabee core architecture is a superwide floating point unit. Capable of handling 16 instructions per clock cycle, it's four times as wide as the equivalent unit in one of Intel's existing desktop CPUs. Factor in each core's additional ability to support four software threads and the chip's potential is truly staggering. It's just possible that Larrabee might deliver 100 times the floating-point punch of a Core 2 Duo chip.

Impressive as that sounds, what we don't know is how good this battery of general purpose and x86-compatible cores will actually be for graphics processing. Sure, it will be much more programmable than any graphics chip before. But that doesn't mean it will be fast.

How it works

The key difference between Larrabee and all previous graphics chips is programmability. The latest DirectX 10 GPUs from AMD and NVIDIA may be much more flexible and programmable than their progenitors, but they retain a wide range of fixed function units for texture sampling, rasterization, geometry setup and more. Only the shader units are truly programmable and even those are much less flexible than a general purpose CPU.

A series of pipes

With Larrabee, the entire rendering pipeline is handled in software by a battery of small x86 cores. The only fixed function units are texture samplers, which Intel says allow for more efficient texture throughput and anisotropic filtering. In effect, therefore, the chip's entire computational resources can be used to attack the most complex task at any given moment in time. In systems with up to 32 cores, benchmarking shows performance increasing in approximately a linear fashion relative to the number of processing cores.

Not only does that programmability make Larrabee theoretically more efficient than previous GPUs, it also means that only a software update will be required to allow Larrabee to support future graphics APIs.