Larrabee has finally died. The project that began as a major GPU initiative within Intel before being shifted to focus on x86 many-core compute has been canceled. Intel has quietly notified customers that the Xeon Phi 7295, 7285, and 7235 will be end-of-life’d July 31, 2020, with no further orders for KML after August 9, 2019. May 6, 2019 marked the beginning of Product Discontinuance Program Support for KML chips.
Since Intel had two similarly named products, let’s briefly discuss. Knights Hill, which was killed off several years ago, was originally intended for 10nm and 2016. Intel killed Knights Hill in November 2017 (its delay pushed back the introduction of the Aurora supercomputer). Knights Mill, in contrast, was a 14nm derivative of Knights Landing (KNL) and designed for AVX-512 workloads, deep learning, and AI acceleration. Unlike previous Larrabee designs, Knights Mill wasn’t a PCIe add-on board, but a socketed chip intended for LGA3647.
While the original Larrabee and Knights Corner were derived from the classic Pentium processor, Knights Mill offered 64 – 72 CPU cores based on the Silvermont x86 Atom architecture, with support for more threads and features like AVX-512. Larrabee architect Tom Forsyth wrote a blog post about his work on the project several years ago that ought to be required reading for anyone interested in the chip. While Larrabee is generally regarded as a failure, he points out that the chip design effort made a major long-term contribution to Intel’s product lines, noting:
The only compromise to full x86 compatibility was KNF/KNC didn’t have SSE because it would have been too much work to build & validate those units (remember we started with the original Pentium chip). KNL adds SSE legacy instructions back in, and so really truly is a proper x86 core. In fact it’s so x86 that x86 grew to include it – the new AVX512 instruction set is Larrabee’s instruction set with some encoding changes and a few tweaks.
Intel isn’t retreating from the HPC space, of course — it’s just planning to address it differently. Where Xeon Phi was an attempt to stretch x86 into a MIC, the company now appears to be focused on Xe, its upcoming graphics architecture, to tackle the AI/DL market. It will take over this role from Xeon Phi, but not until 2020. Whether or not Intel will re-use the branding isn’t known; presumably, the company will create a new brand around Xe given the amount of attention being paid to the new effort. The decision to kill the architecture is supposedly directly tied to the problems surrounding 10nm, which makes sense. If Intel had shipped Knights Hill in 2016, as originally intended, it would already be shipping 7nm today and would have been able to bring new products to market to compete with the likes of Nvidia. Instead, the company has been stuck with 14nm hardware and a limited roadmap to compete against rival firms.
Intel hasn’t really started talking about Xe as a solution for AI and deep learning yet, but we know that these markets are key targets for the GPU — it’s no accident that Intel is moving into the space now, after decades and multiple false starts to do so. The advent of AI and deep learning means having a GPU is now non-optional if you want to play in the data center space — and that market is far too important to Intel not to compete for it. Xe, not Xeon Phi, is the future of Intel’s compute framework for AI and DL — with plenty of assistance from Xeon, of course, where appropriate. Whether Xe will deliver the goods or not is something we’ll have to wait a year to find out.