After a week of media buzz about Intel’s upcoming manycore Larrabee processor, I thought I’d try to get a sense of how the competition — namely AMD and NVIDIA — is reacting to the news. If Intel is able to deliver the goods with Larrabee, both its rivals have a lot to lose.

With Intel in the mix, all three chip vendors are now looking to expand their share of the visual computing and supercomputing pie. Intel, though, is definitely taking a different path. Larrabee’s integration of a stripped-down x86 with a 512-bit SIMD unit gives it a very HPC flavor compared to the standard GPUs offered by its rivals. One way to look at it is that AMD and NVIDIA (and IBM, for that matter) took a game processor and transformed it into an HPC platform, while with Larrabee, Intel is attempting to perform the trick in reverse. As I pointed out in this week’s feature article, from Intel’s point of view, this is a necessary strategy since the company needs to build volume and ecosystem support in the consumer graphics space before venturing further afield.

What Intel is presumably hoping for is that in the next decade Larrabee will be the platform of choice for a new set of terascale applications, which incorporate both visual computing and HPC. The company’s concept of this is something called RMS (Recognition, Mining, and Synthesis), which describes a set of applications that manipulate complex models and events. Until this particular Nirvana arrives, Intel has to beat the GPU vendors at their own game, so to speak. From just a performance point of view, that means Larrabee will have to achieve multiple (single precision) teraflops to match up well with GPUs it will be going up against in the next couple of years.

But Raja Koduri, chief technology officer for AMD’s graphics processing group, doubts Larrabee will be competitive compared to discrete GPUs. “From a performance per watt and performance per dollar perspective, GPUs should still have a big advantage on existing workloads,” he told me. Based on what he could extrapolate from the technical paper that Intel will be presenting at SIGGRAPH later this month, he thought a 16- or 32-core Larrabee chip would be equivalent to the performance AMD offered in their previous generation of low-end GPUs. And according to Koduri, by the end of 2009 or 2010, low-end graphics processors will be three times faster.

Even for general-purpose HPC workloads and user codes such as FFT or matrix multiplication, Koduri thinks GPUs will still have a performance advantage because their compute density will be much higher. He allows that a Larrabee architecture might be useful for software that requires both fine-grained control, which is suitable for multicore CPUs, and highly data-parallel operations, which are applicable to GPUs. But he’s not sure which applications would fit in that sweet spot.

NVIDIA’s take on Larrabee appears to be very similar. Andy Keane, NVIDIA’s general manager of its GPU computing group, says it’s going to be very difficult to build a platform from a set of general-purpose cores that competes against GPU performance or features. From his point of view, the separation of CPU and GPU is still valid since there is no unifying software model that brings those two architectures together.

Keane thinks Larrabee’s dependence on software will limit its utility as a GPU and make it even more problematic to program than a multicore processor. “The multicore scaling problem still exists,” he says, noting that adding more cores just exacerbates the problem. In fact, scaling codes for Larrabee is potentially a lot more challenging than it will be for the current crop of quad-core chips because 16 or 32 cores will be much more difficult to manage.

At this point, I’ll interject that I was a little surprised Intel didn’t at least give a mention to Ct in its first public unveiling of the new architecture. Ct is the company’s parallel programming language being developed for throughput computing. Since Larrabee is a throughput architecture, why not at least give a shout out to the matching language/compiler technology? Even if Ct never makes it out of the lab, I’m sure Intel will end up supporting Larrabee in its vanilla C/C++ compiler and libraries. I can only guess that Intel wants to initially push the Larrabee-as-GPU message, so limited the software talking points to DirectX and OpenGL support.

The other missing piece of the puzzle is the nature of Larrabee’s vector instruction set. Intel kind of glossed over the fact that it was inventing a bunch of new non-x86 instructions, which the user, the compiler, or some other layer of software will have to deal with. It’s been suggested that Larrabee’s vector instructions will be based on Intel’s upcoming AVX vector extensions to SSE, but since the Larrabee unit is 512 bits wide and AVX is currently spec’ed at 256 bits, I’m not sure how this gets resolved. It would probably make the most sense if Larrabee just supported a double-wide version of AVX. Despite Intel’s emphasis on the x86-ness of Larrabee, Keane thinks that the nonstandard vector unit will lock developers into a proprietary architecture.

Intel will have to work hard to get programmers to adopt Larrabee for general-purpose computing since most won’t want to have to deal with assembly code or inline intrinsics. Software companies like RapidMind could help here, since supporting high-level code development for these types of parallel computing platforms is the company’s forte. RapidMind currently supports the Cell processor, GPUs (NVIDIA and AMD), and multicore x86 platforms with its development platform.

Michael McCool, founder and chief scientist at RapidMind, says Larrabee represents a very different design than the existing x86 architecture, GPUs or the Cell processor. “From a software development point of view, it really has to be treated as a new kind of processor,” he explains, adding that software will be the key to extracting performance. “It is important to realize, it’s not necessarily going to be an easy transition from ordinary x86 code to high-performance Larrabee code.”

According to RapidMind CEO Ray DePaul, the company does intend to offer Larrabee support in its product. “This is an important development for Intel and we expect that it will get traction,” he says. “It is a very good target for our platform and we’ll be able to do very interesting things for applications by mapping applications to Larrabee using RapidMind.”

Intel still has 12 to 18 months to refine the architecture and set the software stage before Larrabee’s debut. In the interim, McCool expects the other major players — AMD, NVIDIA, IBM — won’t be sitting still. He believes we’ll see some pretty remarkable evolution of the Cell processor and discrete GPUs over this time period, adding, “I think it will be a very interesting next couple of years.”