A decade later, the company had hoped to make a discrete GPU comeback with Larrabee, a 45nm 32-core GPU which would use x86 instructions. However, the program has been chronically postponed even as advanced new GPU designs from ATI and NVIDIA have come on the market.

Intel has now decided to cancel the consumer GPU version of Larrabee which was supposed to have come out next year. It was supposed to feature two teraFLOPS of performance, but ATI broke that barrier earlier this year.  A teraFLOPs is 1 trillion FLoating point Operations Per Second, and is an indicator of CPU and GPU performance.

AMD's graphics division launched the 40nm Radeon HD 5870 with 2.72 teraFLOPS just before Intel showed off a prototype of Larrabee at the Intel Developer Forum in September. The recent launch of the Radeon HD 5970 with over 5 teraFLOPS was the final nail in the coffin. Intel decided that Larrabee just wouldn't be able to compete on price or performance.

"Larrabee silicon and software development are behind where we hoped to be at this point in the project," stated Intel in a email to DailyTech.

"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".

"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," the statement concluded.

The initial software development platform will be launched next year. The company had previously stated that Larrabee was to be just the first of several GPUs.

Larrabee was supposed to combine the raw parallel throughput of a GPU with the general programming ability of a CPU. Intel often highlighted rendering features that are difficult to achieve on GPUs like real-time raytracing and order-independent transparency. Larrabee would enable features like that through its tile-based rendering approach.

Born out of Intel's many-core Terascale Initiative, Larrabee's hardware is based on the Pentium P54C CPU. It contains vector-processing units to enhance the performance of graphics and video applications. The cores featured short instructional pipelines, with support for Hyper-Threading four execution threads per core, which has its own register set to access memory. A short instructional pipeline allows fast access to L1 cache. All cores on Larrabee share access to a partitioned L2 cache, while cache coherency across all cores would maintain data consistency. Communication between all of the Larrabee cores is through a 1024-bit bidirectional ring bus.

Intel recognized the importance of software and drivers to Larrabee's success, leading to the creating of the Intel Visual Computing Institute at Saarland University in Saarbrücken, Germany. The lab conducts basic and applied research in visual and parallel computing.

The Larrabee roadmap showed a future 48-core version built on a 32nm process, and hinted at more powerful versions built using Intel's 22nm SOC process in 2012. While we may or may not see products based on Larrabee, the cancellation of consumer Larrabee GPUs means that Intel will be able to deploy more resources on improving the 32nm integrated graphics which will be found in next year's Sandy Bridge chips. Nevertheless, this whole mess is an embarrassment for Intel, and a major retreat in their ongoing saga with AMD and NVIDIA.