I wonder what this implies about DP versus SP workloads on HPC clusters. The fact that Knights Mill has half the DP performance (and double the SP performance) when compared to Knights Landing implies that Intel misjudged what HPC customers want/wanted.
No support for Crysis meant nobody could benchmark it to determine the answer to the question, "Can it run Crysis?" so no corporation was going to touch Knights Landing with a ten foot pole.
Deep learning applications generally don't need that high of precision. The improvements in single and half precision are what the deep learning niche wants and that is what they got in Knights Mill.
"The Intel Knights Mill product is essentially being launched to fulfill a promise to a customer base that there would be a product, based on Xeon Phi, that would target deep learning training for a scale-out market."
1) When Intel makes bespoke chips, they don't market them, therefore the x205 chips aren't for any particular customer. 2) "a customer base" means just the opposite of "a certain customer". It means "most customers", i.e. the "base" of a pyramid where the top are niche uses/users and the base represents most uses most of the time.
The PCIe card version only offered marginal density improvements while removing one of the chips advantages: memory capacity via six channels of DDR4. The 16 GB of HMC was plenty for providing a high bandwidth cache but the working data sets are often far larger. Being able to host 384 GB of DDR4 on top of the 16 GB of HMC per node was plenty of room. Various deployments were leveraging four nodes in a 2U chassis which is in-line with how many double wide PCIe cards could fit into a similar chassis. The other difference is that the Xeon Phi nodes could also accept a small single slot PCIe card, providing more IO or another accelerator like a FPGA.
The flip side is that that PCIe cards would have had to leverage external fabric but it could have been Omnipath with the Sky Lake Xeons or ordinary Omnipath PCIe card. Scaling outward would have been easier due to the fewer number of nodes but it made the individual nodes far more expensive. Memory capacity of a dual socket Sky Lake-SP system would have been roughly the same due to higher capacity DIMM support and being able to leverage two DIMMs per channel.
Ultimately the loss of the PCIe based Xeon Phis wasn't a big deal for HPC. As noted in the article, the delays to 10 nm production kill off Knights Hill which was to be the next Xeon Phi chip. It would have made sense for Intel to release a Knights Corner refresh for HPC that leverages the same socket as Sky Lake-SP and its features. It would have been a simple injection that leverages platforms that were already in the pipeline to keep Intel in the game. Greater performance per node would have been handled by dual socket boards. While not major changes in the grand scheme of things, these would have permitted Intel to get a product out and maintain their foot hold in the market. Instead we have Knights Mill which changes up the vector unit for deep learning tuning.
Oddly even though PHI was never intended as graphics card - it possible research for Artic Sound and Raju may have made PHI chips obsolete for new technology coming.
"Oddly even though PHI was never intended as graphics card"
A common misconception. Larrabee is Xeon Phi is Larrabee. The die is identical, the only difference is Larrabee is running the Dx-on-x86 software and the PCB has the DVI header populated. ( http://tomforsyth1000.github.io/blog.wiki.html#[[Why%20didn%27t%20Larrabee%20fail%3F]] link will probably break, Forsyth has square brackets in his URLs fro some reason)
In any case, Intel really did not market it as graphics card - and I have been dealing with graphics card since days that AMD graphics were ATI not intelligent devices - I actually talk on phone with ATI developer almost 30 years ago. Also before NVidia GPU were NVIdia - IE 3dfx Voodoo. which is where I started out with my desktop GPU.
"Larrabee is Xeon Phi is Larrabee. The die is identical, the only difference is Larrabee is running the Dx-on-x86 software and the PCB has the DVI header populated"
While Xeon Phi came directly from Larrabee it is not true that the dies are identical, especially Knights Landing (KNL) Firstly, the core was changed. KNL uses a Silvermont core that I believe has been heavily modified, plus it has AVX-512 which is doing the heavy lifting in HPC. Larrabee (as well as first gen Xeon Phi) used a P5 based core. The fabric tying the cores together has also been significantly changed for KNL.
The sad thing here is that Intel should rule the GPU market. Imagine for a moment Nvidia's 1080Ti GPU fabbed on Intel's p1272 process. It would absolutely destroy anything else on the market by a wide margin. Intel is good at one thing and one thing only, and that's producing silicon circuitry. GPU devices fall directly into Intel's strong suit, but the upper management is too stupid to realize it. They're so concerned with hiring people of color that they couldn't care less about the actual product.
HArd to rule anything when you did not make the base line code for what makes a gpu a gpu most if not all the tech and code required for doing so is owned by AMD and Nvidia.
Intel has x86 with x64 support (courtesy of AMD) a corporation no matter how "mighty they are" is capable of owning everything (especially in the tech world)
Apple does not make its own GPU when they can just buy them at a fraction of the cost vs producing, IBM does not (shell of what they once were) many other companies/corporations do not make their own GPU...I wonder why that is?
is it possible it is "not worth it" when they can just buy them from companies that have sunk massive resources and legal IP protection, I think so.
why did Intel purchase Vega (already built and programmed by AMD) on their new APU model, because Intel is unable to build something that is as fully feature/spec rich as buying something pre-built by AMD (Ngreedia head is way too full of air these days) otherwise all mighty Intel according to you would have had the best of the best GPU years and years ago, BUT they never have had "decent" graphics compared to the gaming and overall capability GPU that AMD or Nvidia produce.
they can produce some that just rely on cpu like "grunt" basic 3d type, which is where all their designs have focused, but, folks also want/need stuff that can do advanced 3d and 2d for computer, graphics etc which Intel has never been very capable at, and even something like their Phi cards while are "beasts" are really slouches compared to the real beasts which are the top class Radeons or Geforce cards lets alone the top models from them.
I'll make a prediction that this is exactly what they are doing. They are going to phase in a GPU in 2020 after they phase out the Phi mid 2019. Likely it will be AI / mining focused.
This. There are certainly workloads -- most significantly weather forecasting -- that Phi suited better than either general-purpose CPUs or GPUs, in term of a) perf/Watt and b) can actually be run. Lots of stuff can't be done on GPGPUs.
I think the major difference is that KNL is a processor, not a co-processor or accelerator. It is an excellent idea, but it just turns out, at least for now, that it cannot beat CPUs in combination with accelerators. If accelerators are to be used anyway, then its cheaper and easier to use general-purpose CPUs.
Not a bad lifetime for a chip that debuted in 2013. It had a good run in HPC too, at its peak it had a 6% system share of the TOP500 (https://www.top500.org/statistics/details/accelfam... beating out every accelerator other than Nvidia GPUs. P100 and V100 quickly assumed total dominance in that space after release though.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
23 Comments
Back to Article
Elstar - Tuesday, July 24, 2018 - link
I wonder what this implies about DP versus SP workloads on HPC clusters. The fact that Knights Mill has half the DP performance (and double the SP performance) when compared to Knights Landing implies that Intel misjudged what HPC customers want/wanted.PeachNCream - Tuesday, July 24, 2018 - link
No support for Crysis meant nobody could benchmark it to determine the answer to the question, "Can it run Crysis?" so no corporation was going to touch Knights Landing with a ten foot pole.HStewart - Tuesday, July 24, 2018 - link
PHI were never consider as gaming card - very limited scientific use.jordanclock - Tuesday, July 24, 2018 - link
Woosh.ImSpartacus - Wednesday, July 25, 2018 - link
Yeah, I don't know what Intel was thinking with that.Do they think that any modern corporation would buy hardware that hasn't been tested on Crysis? The thought is ridiculous.
Kevin G - Tuesday, July 24, 2018 - link
Deep learning applications generally don't need that high of precision. The improvements in single and half precision are what the deep learning niche wants and that is what they got in Knights Mill.lefty2 - Wednesday, July 25, 2018 - link
Knight Mill was dead on arrival. It only exists because Intel made a promise to a certain customer:https://www.servethehome.com/intel-xeon-phi-x205-s...
"The Intel Knights Mill product is essentially being launched to fulfill a promise to a customer base that there would be a product, based on Xeon Phi, that would target deep learning training for a scale-out market."
Elstar - Wednesday, July 25, 2018 - link
1) When Intel makes bespoke chips, they don't market them, therefore the x205 chips aren't for any particular customer.2) "a customer base" means just the opposite of "a certain customer". It means "most customers", i.e. the "base" of a pyramid where the top are niche uses/users and the base represents most uses most of the time.
Kevin G - Tuesday, July 24, 2018 - link
The PCIe card version only offered marginal density improvements while removing one of the chips advantages: memory capacity via six channels of DDR4. The 16 GB of HMC was plenty for providing a high bandwidth cache but the working data sets are often far larger. Being able to host 384 GB of DDR4 on top of the 16 GB of HMC per node was plenty of room. Various deployments were leveraging four nodes in a 2U chassis which is in-line with how many double wide PCIe cards could fit into a similar chassis. The other difference is that the Xeon Phi nodes could also accept a small single slot PCIe card, providing more IO or another accelerator like a FPGA.The flip side is that that PCIe cards would have had to leverage external fabric but it could have been Omnipath with the Sky Lake Xeons or ordinary Omnipath PCIe card. Scaling outward would have been easier due to the fewer number of nodes but it made the individual nodes far more expensive. Memory capacity of a dual socket Sky Lake-SP system would have been roughly the same due to higher capacity DIMM support and being able to leverage two DIMMs per channel.
Ultimately the loss of the PCIe based Xeon Phis wasn't a big deal for HPC. As noted in the article, the delays to 10 nm production kill off Knights Hill which was to be the next Xeon Phi chip. It would have made sense for Intel to release a Knights Corner refresh for HPC that leverages the same socket as Sky Lake-SP and its features. It would have been a simple injection that leverages platforms that were already in the pipeline to keep Intel in the game. Greater performance per node would have been handled by dual socket boards. While not major changes in the grand scheme of things, these would have permitted Intel to get a product out and maintain their foot hold in the market. Instead we have Knights Mill which changes up the vector unit for deep learning tuning.
babadivad - Tuesday, July 24, 2018 - link
Larabee's baby.HStewart - Tuesday, July 24, 2018 - link
"Larabee's baby"Oddly even though PHI was never intended as graphics card - it possible research for Artic Sound and Raju may have made PHI chips obsolete for new technology coming.
edzieba - Wednesday, July 25, 2018 - link
"Oddly even though PHI was never intended as graphics card"A common misconception. Larrabee is Xeon Phi is Larrabee. The die is identical, the only difference is Larrabee is running the Dx-on-x86 software and the PCB has the DVI header populated. ( http://tomforsyth1000.github.io/blog.wiki.html#[[Why%20didn%27t%20Larrabee%20fail%3F]] link will probably break, Forsyth has square brackets in his URLs fro some reason)
HStewart - Wednesday, July 25, 2018 - link
In any case, Intel really did not market it as graphics card - and I have been dealing with graphics card since days that AMD graphics were ATI not intelligent devices - I actually talk on phone with ATI developer almost 30 years ago. Also before NVidia GPU were NVIdia - IE 3dfx Voodoo. which is where I started out with my desktop GPU.Yojimbo - Wednesday, July 25, 2018 - link
"Larrabee is Xeon Phi is Larrabee. The die is identical, the only difference is Larrabee is running the Dx-on-x86 software and the PCB has the DVI header populated"While Xeon Phi came directly from Larrabee it is not true that the dies are identical, especially Knights Landing (KNL) Firstly, the core was changed. KNL uses a Silvermont core that I believe has been heavily modified, plus it has AVX-512 which is doing the heavy lifting in HPC. Larrabee (as well as first gen Xeon Phi) used a P5 based core. The fabric tying the cores together has also been significantly changed for KNL.
JKflipflop98 - Wednesday, July 25, 2018 - link
The sad thing here is that Intel should rule the GPU market. Imagine for a moment Nvidia's 1080Ti GPU fabbed on Intel's p1272 process. It would absolutely destroy anything else on the market by a wide margin. Intel is good at one thing and one thing only, and that's producing silicon circuitry. GPU devices fall directly into Intel's strong suit, but the upper management is too stupid to realize it. They're so concerned with hiring people of color that they couldn't care less about the actual product.What a bunch of morons.
yeeeeman - Wednesday, July 25, 2018 - link
They were good, but they hit a wall at 10nm.Dragonstongue - Wednesday, July 25, 2018 - link
Intel should rule GPU market.. AH HA HA HA HA.HArd to rule anything when you did not make the base line code for what makes a gpu a gpu
most if not all the tech and code required for doing so is owned by AMD and Nvidia.
Intel has x86 with x64 support (courtesy of AMD) a corporation no matter how "mighty they are" is capable of owning everything (especially in the tech world)
Apple does not make its own GPU when they can just buy them at a fraction of the cost vs producing, IBM does not (shell of what they once were) many other companies/corporations do not make their own GPU...I wonder why that is?
is it possible it is "not worth it" when they can just buy them from companies that have sunk massive resources and legal IP protection, I think so.
why did Intel purchase Vega (already built and programmed by AMD) on their new APU model, because Intel is unable to build something that is as fully feature/spec rich as buying something pre-built by AMD (Ngreedia head is way too full of air these days) otherwise all mighty Intel according to you would have had the best of the best GPU years and years ago, BUT they never have had "decent" graphics compared to the gaming and overall capability GPU that AMD or Nvidia produce.
they can produce some that just rely on cpu like "grunt" basic 3d type, which is where all their designs have focused, but, folks also want/need stuff that can do advanced 3d and 2d for computer, graphics etc which Intel has never been very capable at, and even something like their Phi cards while are "beasts" are really slouches compared to the real beasts which are the top class Radeons or Geforce cards lets alone the top models from them.
FreckledTrout - Wednesday, July 25, 2018 - link
I'll make a prediction that this is exactly what they are doing. They are going to phase in a GPU in 2020 after they phase out the Phi mid 2019. Likely it will be AI / mining focused.Machinus - Wednesday, July 25, 2018 - link
The KNL PCI cards were excellent products. They are an entire cluster in one card. It is a shame that Intel did not continue to develop these.AMD excels in power-efficient, many-core CPU products. I would like to see them produce a card like this for HPC use.
Meteor2 - Wednesday, July 25, 2018 - link
This. There are certainly workloads -- most significantly weather forecasting -- that Phi suited better than either general-purpose CPUs or GPUs, in term of a) perf/Watt and b) can actually be run. Lots of stuff can't be done on GPGPUs.I guess it just wasn't a big enough niche.
Jaybus - Saturday, July 28, 2018 - link
I think the major difference is that KNL is a processor, not a co-processor or accelerator. It is an excellent idea, but it just turns out, at least for now, that it cannot beat CPUs in combination with accelerators. If accelerators are to be used anyway, then its cheaper and easier to use general-purpose CPUs.iwod - Wednesday, July 25, 2018 - link
So Knights Mill is the Kabylake or Coffeelake of Skylake. What exactly is the big deal EOLing an earlier processors? Am I missing something obvious?edzieba - Wednesday, July 25, 2018 - link
Not a bad lifetime for a chip that debuted in 2013. It had a good run in HPC too, at its peak it had a 6% system share of the TOP500 (https://www.top500.org/statistics/details/accelfam... beating out every accelerator other than Nvidia GPUs. P100 and V100 quickly assumed total dominance in that space after release though.