Discussing the next processor step after Penryn and entry into the discrete graphics market.
Intel’s Tick-Tock strategy has been discussed before before, in reference to the launch of Penryn, and by now most should be familiar it. That particular release was a die-shrinking ‘Tick’ and now the excitement over 45nm has died down, Intel has laid out some more details about the forthcoming ‘Tock’, known as Nehalem, and its more distant cousin, Larrabee. Of the former we’ve already heard about as much as Intel is ready to say, but the latter has been something of a mystery – in both cases our curiosities have been yet further aroused.
Kicking off with Nehalem we have a reaffirmation of such features as Simultaneous Multi-Threading (SMT), basically a rework of HyperThreading, the modular nature of the Nehalem processor, whereby individual cores and their cache allocations can be dynamically switched on and off and, the most radical change, the move to an integrated memory controller. Using DDR3, this memory controller will also offer support for up to three memory channels, improving bandwidth and latency. The latter of these is a big problem with current DDR3 systems, as often the benefit of DDR3’s increased speed is mooted by its higher latencies.
The modular nature of the CPU also means that various features can be added and taken away depending on the intended user-base. These include cores, integrated memory chips and QuickPath interconnects – the 25.6Gbit/s channel over which cores communicate with each other. Basically, the higher-end the CPUs intended user, the more ‘components’ can be added. From the early figures available Nehalem looks set to absolutely wipe the floor with AMD’s current processors. We’ll find out near the end of this year when Nehalem launches.
Next up, and arguably more interesting, is Larrabee, marking Intel’s entry into the discrete graphics market. Unlike AMD, though, which did so by purchasing an existing GPU maker, Intel has done things the hard (or at least long) way and created a graphics architecture all of its own from the ground up.
Unlike current cards from AMD/ATi and nVidia, Larabee will run using the x86 instruction set implemented on Intel and AMD’s current CPUs – albeit a modified version. The GPU will comprise a set of Intel Architecture (IA) cores which won’t compare, in terms of raw processing power, to a desktop part, but will be optimised for the more specific functionality it is intended to provide so don’t expect it to be sluggish either.
Importantly the DirectX and OpenGL APIs will be supported, so current generation (and older for that matter) games will all function on Larrabee. Of course, as well as gaming, a variety of other tasks will also see potential for offloading, much like with nVidia’s CUDA program and AMD/ATi’s CTM. 3D rendering and computational modelling are two examples given that would benefit from running on Larrabee – there is even talk in some circles that real-time ray tracing might be a possibility, assuming there is software capable of using any such potential available.
Larrabee is currently scheduled for release late 2009 or early 2010 so we’re not likely to see anything physical for a while yet. nVidia and AMD should have new architectures ready by that time to offer some competition, so it’s hard to predict how Larabee will fare. One thing is for sure: we’re in for an interesting battle sooner or later.