Xeon Phi
Intel's x86 manycore coprocessor family for supercomputers and HPC
Xeon Phi was a series of x86 manycore processors designed and made by Intel, launched in 2010 and discontinued in 2020. Built for supercomputers, servers and high-end workstations, it ran standard parallel programming tools such as OpenMP and competed with GPGPU products like Nvidia Tesla.
- Launched 2010, discontinued 2020; 32 to 72 x86 cores per chip
- Tianhe-2 reached 33.86 petaFLOPS with Xeon Phi coprocessors
- 512-bit SIMD vector units, later aligned with AVX-512
- Knights Landing booted as a standalone CPU in an LGA 3647 socket
- Knights Hill, the 10 nm generation, was cancelled in 2017
What it is
Xeon Phi was a family of x86 manycore processors built by Intel between 2010 and 2020. It was aimed at supercomputers, servers and high-end workstations, and it handled heavy parallel computing tasks. Each chip carried dozens of small x86 cores, so ordinary parallel programming tools such as OpenMP, OpenCL and Intel's own Fortran and C++ libraries could be used with it. That set it apart from a GPGPU like Nvidia Tesla: because the cores spoke x86, existing x86 software could be moved across with fewer changes. The line competed with Nvidia Tesla and AMD Radeon Instinct for machine learning and general-purpose GPU-style work.
How it came about
Xeon Phi grew out of Intel's Larrabee project, a GPU design started in 2006 and cancelled as a retail product in May 2010. The first board, Knights Ferry, was announced on 31 May 2010 as a Many Integrated Core prototype. Early users included CERN, KISTI and the Leibniz Supercomputing Centre. The name Xeon Phi was adopted in June 2012 at the Hamburg supercomputing conference for all Many Integrated Core products. Knights Landing, the second generation, was revealed in June 2013 and launched as the x200 family in June 2016. The line was discontinued due to weak demand and Intel's difficulties with its 10 nm node.
Inside the chip
Each core is a small x86 design derived from the P54C core used in the original Pentium, running four threads at once. Cores are linked by a very wide ring bus to memory and share a coherent L2 cache of 512 KB per core, with 32 KB of L1 cache each. The heart of the chip is a vector processing unit with 512-bit SIMD instructions, able to handle 16 single-precision or 8 double-precision operations per cycle, and twice that with fused multiply-add. An Extended Math Unit speeds up reciprocals, square roots and logarithms. Knights Landing added two 512-bit units per core and AVX-512 support.
How it was made
The family spanned several Intel process nodes, from 45 nm on the Knights Ferry prototype to 22 nm with tri-gate transistors on Knights Corner, and 14 nm on Knights Landing. Chips were built by Intel itself. Knights Landing sat in an LGA 3647 socket or, in earlier cards, on a PCI Express 3.0 x16 board. It paired up to 384 GB of DDR4 memory with 8 to 16 GB of stacked MCDRAM on the package for much higher bandwidth.
Where it went
Xeon Phi went into some of the world's largest supercomputers. Tianhe-2 at the National Supercomputer Center in Guangzhou combined Xeon Phi coprocessors with Ivy Bridge-EP Xeon E5 v2 chips to reach 33.86 petaFLOPS, topping the TOP500 list from June 2013 until November 2015. TACC's Stampede system used Knights Corner cards for 8 petaFLOPS of its 10-petaFLOPS total. NERSC's Cori machine chose Knights Landing for its second phase. Cray offered the chips in its Cascade systems, and IBM, SGI, HP and Dell built prototype hardware.
Versions and family
Knights Ferry was the 2010 prototype with 32 in-order cores at up to 1.2 GHz and 2 GB of GDDR5. Knights Corner, at 22 nm, was the first commercial product, sold as the 3100, 5110P and 7120P. Knights Landing, the x200 family, offered up to 72 Airmont cores and could boot as a standalone CPU; some models carried an integrated Omni-Path fabric, marked with an F suffix. Knights Mill, released in December 2017, retuned the design for deep learning. Knights Hill, the planned 10 nm third generation, was cancelled in 2017.
Why it matters
Xeon Phi put manycore x86 hardware into several of the fastest computers on Earth and helped keep Tianhe-2 at number one for two and a half years. It showed that wide vector units and dozens of small cores could deliver over a teraFLOP of double-precision work from a single chip. Its 512-bit instruction set shared many functions with AVX-512, which went on to appear in later Intel CPUs. Studies found real speed-ups in life sciences and deep learning when the thread and vector parallelism was used well.
Problems and reception
Work on Xeon Phi showed that high performance still needed skilled programmers; compilers and traditional models alone were not enough. The PCIe coprocessor version of Knights Landing never reached the general market and was dropped by August 2017, and the whole Knights Landing line was discontinued in summer 2018. The product family ended in 2020, blamed on weak demand and Intel's 10 nm problems.




Questions people ask about Xeon Phi
What made Xeon Phi different from a graphics processor?
Its cores were x86-compatible, so software written for ordinary x86 CPUs could run on it with fewer changes. A GPGPU such as Nvidia Tesla usually needs code rewritten for its own model, while Xeon Phi worked with tools like OpenMP and standard Intel compilers.
How many cores did these chips have?
Across the family the count ranged from 32 to 72 cores. The Knights Ferry prototype had 32 in-order cores, Knights Corner carried more than 50, and Knights Landing offered up to 72 Airmont cores with four threads each.
Could Xeon Phi work as a normal CPU?
The first generation was a PCI Express add-in card only. From Knights Landing onward the chip could boot as a standalone processor in an LGA 3647 socket, though a PCIe coprocessor version was planned and later dropped.
Why was the family discontinued?
Intel ended the line because demand was weak and the company ran into trouble with its 10 nm manufacturing process. Knights Hill, the planned 10 nm successor, was cancelled in 2017 and the last products shipped around 2020.
What memory did a Xeon Phi system use?
Knights Landing supported up to 384 GB of DDR4 2133 memory plus 8 to 16 GB of stacked MCDRAM on the package, giving over 400 GB/s of bandwidth from the near memory and up to 115.4 GB/s from DDR4 with ECC.
Good reading and sources
- Some Surprises in the 2018 DoE Budget for SupercomputingThe Next Platform, Article
- Checking your browserAnandTech, Article
The facts on this page come from the Wikipedia article "Xeon Phi", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryChip rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.
Like Xeon Phi
All CPUs and microprocessors
🇺🇸 United StatesLarrabee
Cancelled Intel x86-based GPGPU project that became the Xeon Phi lineGPU · x86 · 2009