The bulk of microprocessor giant Intel’s (NASDAQ:INTC) Data Center Group (DCG) revenue comes from the sale of its Xeon family of general purpose processors. With each generation, these chips become both more powerful and more power-efficient.
Ultimately, Intel hopes that those improvements in performance and power efficiency lead to a compelling enough reduction in the total cost of ownership for its major customers that they are compelled to upgrade to the company’s latest chips.
For most data center workloads, better general-purpose processors are still the way to go, as evidenced by Intel's large and growing data center business. Over the years, though, an increasing number of performance-critical data center workloads have been adapted to run highly parallel graphics processors sold primarily by NVIDIA (NASDAQ:NVDA).
Not all workloads are suitable for these highly parallel processors, but the ones that are often run much better on those specialized processors than on general-purpose CPUs.
In a bid to keep customers from moving from Intel’s general-purpose Xeon processors to NVIDIA’s graphics processors, Intel has invested in its own line of highly parallel processors, marketed as Xeon Phi, to compete primarily with NVIDIA’s graphics processor technology and blossoming software ecosystem.
Intel’s first commercially viable version of Xeon Phi, known as Knights Corner, launched in 2013, and a next-generation version, known as Knights Landing, hit the market in 2016.
The company had previously indicated that it intended to launch a successor to Knights Landing, known as Knights Hill, in 2018. In fact, the U.S. Department of Energy had announced that it planned to use these Knights Hill chips in the Aurora supercomputer at its Argonne Lab, which was expected to be operational in 2018.
It would seem, though, that Knights Hill has been canceled.
Per a slide presented at the International High Performance Computing Forum (a Twitter user with the handle HPC Guru tweeted a picture of the slide), the U.S. Department of Energy now expects to get ahold of an Intel-based supercomputer for the Argonne Lab in 2021. But that slide explicitly said that the supercomputer will be based on upcoming Intel processors different from Knights Hill.
Moreover, the slide doesn’t explicitly indicate that it’ll be a future Xeon Phi part, either.
Though an Intel spokesperson told me that “Intel Xeon Phi processors continue to be a key element of [Intel’s] solution portfolio for [high performance computing] and artificial intelligence,” I don’t think the delay and shift in technical direction for the Aurora supercomputer bodes well for Knights Hill or its direct successors.
Based on this, it seems reasonable to conclude that Knights Hill and any future chips based on the same basic technical direction have been canceled.
Impact to Intel’s business
In the near term, I don’t think the cancellation of Knights Hill or even its successors will have a noticeable impact on Intel’s business. Though Intel doesn’t break out the size of its Xeon Phi revenue, there is evidence that it’s an incredibly small part of the company’s data center group business (for more details, see this analysis).
However, over the long term, I think Intel will need a viable competitor to NVIDIA’s graphics processor for highly parallel workloads. By not fielding competitive products, Intel not only misses out on the incremental revenue opportunity that the shift toward parallel computing brings, but it will be powerless to stem the cannibalization of Intel’s bread-and-butter Xeon processors by NVIDIA’s general-purpose graphics processors.
That said, if the Xeon Phi line really has come to an end, this may be for the best. With Xeon Phi, Intel was never able to match the performance of contemporary NVIDIA graphics processors, and the gap seemed to only be growing wider.
If, from the ashes of Xeon Phi, a newer, more competitive approach to highly parallel computing emerges, then that’d put Intel in a far better position over the long term than simply giving up or continuing to field relatively mediocre parts would.