While Intel prides itself on maintaining a breakneck speed for processor development, the company’s Larrabee GPU effort just couldn’t to keep pace with graphics technology development at rivals NVIDIA and AMD. Intel revealed late last Friday that the company would not be delivering a Larrabee-based discrete graphics product next year, and has instead decided to use the work as the basis for a software development platform.

The company blamed the cancellation of the product release on hardware and software schedule slips, preventing Intel from launching its standalone, x86-based GPU offering in 2010. In essence, the chipmaker has withdrawn Larrabee from its commercial plans and recast it as a research project.

The statement from Intel reads as follows:

Larrabee silicon and software development are behind where we had hoped to be at this point in the project. As a result, our first Larrabee product will not be launched as a standalone discrete graphics product, but rather be used as a software development platform for internal and external use.

Larrabee is a large complex, multi year project that is aimed at creating a new architectural approach to graphics and to high performance computing.

The performance of the initial Larrabee product for throughput computing applications — as demonstrated at SC09 — is extremely promising and we will be adding a throughput computing development platform based on Larrabee, too.

While we are disappointed that the product is not yet where we expected, we remain committed to delivering world-class many-core graphics products to our customers. Additional plans for discrete graphics products will be discussed some time in 2010.

Our plans to deliver the world’s first CPU’s with integrated graphics this month are unchanged.

The implication is that the R&D effort might eventually yield commercial offerings, whether or not they are released under the Larrabee moniker. For the time being though, AMD and NVIDIA have received an early Christmas present. For the next year, at least, the two GPU makers will have the market to themselves, both in traditional high-end graphics and for GPU computing. Both companies are fielding teraflop-level graphics processors today, with performance and general-purpose capabilities continuing to spiral upward.

Perhaps the most interesting part of the announcement is how wide Intel left the door open to using the Larrabee work as a springboard to HPC-type applications. Up until now, the official story line for the GPU wannabe was that it would be introduced as a discrete graphics chip, and its HPC duties would be limited to experimental use. With the discrete graphics line on hold, Intel now seems more interested in exploring its high performance potential.

The reference to “adding a throughput computing development platform based on Larrabee” could mean that Intel has decided to formally link the Ct research language and Larrabee development together. Ct, which stands for “C/C++ for throughput computing,” is a new parallel programming language that the chipmaker is developing for manycore processing. While Ct is meant to be a general-purpose, data-parallel programming language, it always seemed to be particularly well-suited to the Larrabee architecture.

The Larrabee demo at SC09 last month suggests Intel was already in the process of expanding its thinking about how the technology can be applied. At the conference, Intel CTO Justin Rattner pushed the idea that the 3D Web, a cloud platform that encompasses real-time simulations, multi-view animation, and immersive virtual environments, would be the application driver for HPC in the next decade. The climax of his presentation was a 1 teraflop Larrabee performance running SGEMM (Single-precision General Matrix Multiply). If Intel now believes a Larrabee-type platform has a bigger future in server-side computing, rather than client-side computing, that points to a telling shift in its GPGPU strategy.

The problem for Intel is that its GPGPU rivals already have commercial offerings in the field — AMD with FireStream, and NVIDIA with Tesla. NVIDIA, in particular, has established a broad ecosystem for GPU computing based on its CUDA hardware/software architecture. While the CUDA software environment does not offer the high-level programming environment of Ct, a broad spectrum of ISVs and tool developers have already coalesced around CUDA. In 2010, NVIDIA is set to release its most advanced and most general-purpose Tesla product line for HPC, based on the Fermi architecture. The company also just started shipping its RealityServer platform for 3D Web applications. While the RealityServer doesn’t support the real-time animations/simulations envisioned by Rattner’s 3D Web, it’s certainly a step toward that direction.

It’s possible that Intel will decide to split its discrete graphics processor effort from its “throughput computing” work and simply develop an architecture for each one. On the other hand, the company may just regroup and develop Larrabee 2.0 against the new realities of the GPU market. In the meantime, GPU computing is back to a two-horse race.