formally announced today that its controversial and much-hyped Larrabee
GPU will not hit retail stores in 2010. The company declined to speculate on when retail Larrabee-based GPUs will be available, but stated that the current generation of products will be used for in-house development and sampled to relevant partners. Intel declined to give a specific reason for the delay, saying only that product development was behind where the company had hoped it would be by now. The project's current status is now a bit unclear, but we know Intel hasn't completely killed the project. Instead, software (and possibly hardware) development kits will be made available over the next year to interested developers.
The news boosted both AMD
and NVIDIA's stock price; NVIDIA
closed at $16.09 (up 12.8 percent) while AMD closed at $8.52, up 8.4 percent. That's good news for both companies in the short-term, but Intel's decision to delay Larabee's retail launch will have no effect on the product's longer-term competitive potential, provided it has one. Graphics and GPGPU capabilities are an increasingly important part of modern computing, and therefore of significant interest to Intel. . Larabee, for those of you who might not be aware of it, is Intel's first discrete graphics processor since the ill-fated i740 launched and promptly sank in 1998. (Historical tidbit: Intel adapted the i740 into an integrated GPU, making it a distant ancestor to the modern X4500HD.)
This diagram is authoritative proof that Larrabee= Win
The controversy around Larrabee has focused on both the architecture's design and Intel's comments regarding the future of 3D graphics. Architecturally, Larrabee is x86-compatible, modular, and (according to Intel), more flexible than a traditional GPU. Much of the work that's handled in hardware by current GPUs would instead be handled in software; a fact that theoretically allows the GPU to dedicate more of its computational resources where they are needed most.
AMD and NVIDIA have both taken swipes at Larrabee's design and performance, but today's announcement is less a relief than it might seem. Intel has the financial resources to develop Larrabee until the architecture is either ready for launch or ready for the scrap heap. As far as its existing product lines are concerned, Intel doesn't need
Larrabee. NVIDIA and AMD, in contrast, very much need each architectural refresh to succeed—when they don't, ala the HD 2000 series, it hits both companies' bottom lines hard.
Larrabee's architecture, in a block-level flavor. All GPU caches are fully coherent, unlike existing Radeon or GeForce products
The counter to the "Intel Juggernaut > All" point of view is that even Intel sometimes bites off more than it can chew. As we've previously noted, the original i740 was a flop, RDRAM never took off, and Itanium utterly failed to meet the original goals Intel had for the platform. (Jump back to 1998, pre-Merced launch, and you'll find pundits prophesying that we'd all be using IA-64 by now. In reality, not so much.) Building a GPU architecture around x86 allows Intel to leverage decades of expertise, but NVIDIA has spent sevearl years already pushing its own solutions, and will have spent at least an additional year before Larrabee hits the market. Waiting to launch gives Intel more time to polish the product, but it also raises the performance bar Larrabee will need to clear. We'll have to wait until 2011 to see if delaying Larrabee was the right call.