Intel Xeon Phi
Images are sometimes not shown due to bandwidth/network limitations. Refreshing the page usually helps.
You are currently reading a thread in /g/ - Technology
You are currently reading a thread in /g/ - Technology
Hey fa/g/s,
Has anyone here been working with intel xeon phis? If yes what where you doing with it?
What its purpose?
>>52994963
to coprocess
>>52994963
x86-64 manycore for parallel computing instead of shitty proprietary closed gpus.
Finding the meaning of life and all that exists
>>52994995
How does it compare to amvidia cards in paraller stuff?
>silvermont atom cores
>>52995075
It looks like it still need some catch up to do. But being actual x86 cores may be an advantage for code more complicated than the usual matrix crunching.
>atom cores
Is this a joke?
>>52995093
>>52995225
>i don't know what i'm talking about LOL!xD
>>52995225
>>52995093
>72 modern atom cores
>bad
The card in OP's pic doesn't even use them, it's an older model based on shrunk Pentium 1 cores.
>>52995284
Thank you pajeet, $0.75 have been deposited to your account.
>atom cores
FUCKING DROPPED!
>~$3,000 price tag for intel's 61-core 1.2GHz xeon phi
>atom cores
ebbin
>>52995540
Do you really think they would be able to fit 61 non-atom cores on one card under one heatsink without the thing being an instant housefire?
>>52995561
Many Core-Ms have no active cooling and they work just fine even above 1GHz frequencies. Xeons Phis have active cooling so cooling 61 @ 1.2GHz skylel or even haslel cores isn't that unrealistic desu.
I got all excited until I saw they used atom cores. Almost got me there jewtel.
>>52995765
if atom cores halted your excitement, then really, you had no use for it to begin with
>>52995540
you get it almost for free nowadays. they are on sale everywhere. Intel is preparing for next generation.
e.g. http://www.thinkmate.com/system/gpxi-xt24-2460-8phi
>>52995804
Would have been an interesting co-processor to use for ray tracing. Too bad they barely have better performance than an1,000 dollar i7-5960x (8 haswell cores).
>atom
>>52995932
Are you talking about single unit or machine with bunch of cards with equivalent cost?
ITT: People compare a coprocessor to the processor in their gaming rigs and fail to realize that largely parallel beats raw power in some applications
>b-but Atom cores
Leave.
>>52996038
A bunch of cards I guess. I'm very interested in real-time (at least 30FPS) accurate ray tracing in 1080p or higher resolutions.
>>52996180
I can't wait for Imagination to get in the game with PowerVR: http://www.tomshardware.com/news/imagination-real-time-ray-tracing,30932.html
A fucking 10w card with passive cooling.
Until then I think GPU-based acceleration is the way to go.
>>52994945
Is this the remains of Larabe?
>>52996835
Larrabee was different, all that stuck was the concept of using a multitude of x86 cores for performing parallel tasks, but I guess you could say that it inspired what eventually became Phi. x86 just sucks for rendering graphics.
>>52996835
Yup
>>52995225
Decent choice for applications like web serving. Don't need much singlethreaded performance but need tons of concurrent performance.
>>52996989
What makes a xeon phi co-processor a better alternative to a 2.3-3.6 GHz16-core xeon CPU with 32 threads, a TDP of 135 watts, and with access to quad channel 2133mhz ddr3 RAM?
>>52994945
These run Linux don't they? iirc you can log in and work right on the device, by any chance do they provide something like a serial port interface over PCIe or do you need drivers to do anything?
>>52997130
>alternative
No
Phi is a expansion card, you would add 60 x86 cores to the xeons you already have
>>52994945
>retards spouting it's Atom based so it's bad
>meanwhile some of the most powerful supercomputers in the world utilize tens of thousands of these
Would this matter for CPU based encoding?
Some people still swear CPUs instead of GPUs in the video editing industry.
>All these idiots complaining about atom cores.
Fucking idiots who don't know how GPU's work.
Atom cores are actually pretty good for this.
>>52994945
i was employed as a student researcher when they were coming up and ran tests on one to see how to work with them and get the best out of them. i had one of the smaller ones, 58 cores i think, mounted in a system on our uni network.
what do you want to know?
>>52997269
Are they seen as CPU cores in Windows?
It know this would be a dumb thing to do.
>>52997269
How do they compare to regular gpus or few powerful cpus? What is the best suitable task for them?
>>52997269
Those extra cores do not just show up along side your system processor cores and are not directly visible to the host OS!
nvm
>>52997294
intel afaik has no driver support, nor sdk support, nor do they have any kind of meaningful high performance computing application support for windows. and intel dev doing a seminar on intel c compiler (icc, which by the way mops the fucking floor with gcc on intel cpus) commented on windows compatibility by saying "irrelevant for the hpc market anyway" or something along those lines.
so no, to my knowledge xeon phis aren't recognized at all under windows.
>>52997310
in raw power, mine wasn't as good as an nvidia tesla, for example, but i couldn't quantify that since i never ran a benchmark against that. i did compare to the xeons of the host system, compared to which the single card was still slower. DGEMM reaches 837GFLOP/s if that tells you anything. the main selling proposition of the card is the fact that is has a ultra fast ring bus for the cores and each core has a VPU that can do 512bit floating point operations, which is what sets it apart from your standard atom cores. they're quite fast when you have large parallel jobs and thereby free the capacity of the CPU to do horse power operations.
as for gpu comparison: nvidia CUDA is a huge pain in the ass if you have to convert your application into it. working with xeon phi and icc is much easier and much more fun; you simply spick your C with a bunch of OpenMP and a few intel tricks and that usually does the larger part of the work. you CAN get down to assembly level but intel's goal is that you don't have to, and they're doing a very decent job at it too.
>>52997317
what now? i never said that... my card had 58 cores, the host cpu was an 8 core xeon.
that being said, you can write the code in such a way that when compiled with icc it looks for a xeon phi and if it finds one outsources the computation there.
>>52997606
Thanks
>>52997684
you're welcome. if you have any more detailed questions, i have my material here.
>>52997749
so basically to sum it up:
slower than cuda but much easier to work with
>>52997749
What kind of projects were you working on with it?
>>52995225
>>52995093
>>52995540
>HEY, WHY ISN'T INTEL PUTTING 100 SIX CORE i9s ??? I BET IT WOULD FLY
Retards.
>>52994945
oh wow real technology
i did some research in school on it. i rewrote our framework that was used to do GPU offloading onto the coprocessor. this was just around 3 years ago though so i'm a bit out of the loop.
it was a real blast to get shit to work on it. the documentation was terrible and the tooling was soso, but pretty satisfying.
result was a small paper comparing task execution via OpenCL vs. framework and offloading to the coprocessor vs. a many machine (24 cores?).
>>52997772
well, technically it's the hardware that's slower. at least as of knight's corner. when they release knight's corner, that might be a different story.
if you write for CUDA directly, CUDA is also fine, but if you're translating your code to it it sucks. programming for xeon phi is basically writing parallel C code with the help of OpenMP and adding a bunch of intel pragmas as well.
>>52997788
running benchmarks, writing benchmarks, running benchmarks, writing benchmarks...
you do mainly calculation heavy benchmarking though, nothing like b-tree sorting or stuff like that. think calculating pi, multiplying matrices, mandelbrot etc. and of course adapting other people's benchmarks and running those.
>>52997939
ops openmp
>>52997917
also, they're not atom cores. they're similar, but with quite some extra oomph in terms of low latency cache and the 512bit VPU.
the upcoming knight's landing architecture is supposed to get something along the lines of 72 quadthread atoms with the extended avx512 vpu attached, douple knight's corners bus width and on-die memory etc. it will fly.
>>52997939
>>52997979
similar to what i did
If I find one cheap, let's say 200$ should I grab it? Will it be worth the money? Or quadro is better option?
>>52998051
did you ever do any work with multiple xeon phi nodes?
the extension of my work was to get a framework running on TACC on top of an existing scheduling system. never got to it because I graduated and now suck the sweet corporate dick.
>>52998186
no, my job was to explore the programming and performance capabilities for the higher up faculty members so they knew what to expect when working with our 3840 phi core cluster extension. i worked with the host system of that during my hpc coursework, but never during my research.
>>52997939
tl;dr on the paper? Any plus sides and negative side not already discussed?
>>52998433
(at the time, not sure if this is relevant), on short running tasks that are sent out individually (<10ms) using our framework (directly push jobs over SCIF) we reduce overhead of OpenMP. batched jobs achieve essentially the same overhead (although it's really a very different job type)
Can I use this card to render 4k videos in Adobe Premiere?
>>52998027
the cores are extremely stripped down, they lack basic optimizations like out of order execution. the only strength in xeon phi is the vector processor, it's pretty much a GPU without all the extra useless fixed function hardware like video decoders/encoders, ROPs, texture compressor/decompressors that would be in the typical pooinloo or nvidia gpu
>>52994945
honestly all your going to get here is a bunch of dumb bitches trying to compare it to the pc they just learned how to build. You in the wrong place nigga.
>silvermont cores
>>52999745
xeon server CPUs > xeon phi memes for a server you dumb motherfucker.
We don't want these pieces of shit either.
>>53000000
>>53000017
/g/ is such a bad board on all lvl's. I dont understand what you are trying to say all I said was people here wont even understand what this is for and compare compare apples to oranges.
>>52996043
>some applications
But any important ones?
>>53000350
the most important ones but you would have to want to very complex calculations.
>>53000643
want to do*
Heptapentaconta-core
>>52994945
no fans?
>>52994945
phi makes my dick hard
>you will never do high-performance scientific computing
>>52999745
plus it's prasident's day so all the kids are home
>>53000934
I'm a wannabe scientist too. one day maybe.
I'm thinking of getting one because I've always been intrigued by them and I have some money to burn. You can get one of them for about 250 to 300 euro on eBay now so I'd be fine with it.
What's something fun I can do with them that could maybe help me learn something in the process?
>>52994945
Some shitty weather modelling app. You can use them in opencl apps with little effort.
>>52999450
>it's pretty much a GPU without all the extra useless fixed function hardware like video decoders/encoders, ROPs, texture compressor/decompressors that would be in the typical pooinloo or nvidia gpu
Soooo...like an Nvidia Tesla?
>>52994945
They're really annoying to program things for.
The APIs for it are almost non existent and it really isnt a good alternative to GPGPU.
>>52994945
Mining buttcoins.
>>52994945
Large scale NodeJS deployments :^)
>>52995093
>>52995225
>>52995344
>>52995376
>>52995540
>>52995765
>>52996033
>>53000008
absolute fucking idiots. Each core has 128 x 512-bit vector registers. Binary operations takes two cycles, arithmetic 4, convert 5, and permute/extended math 6 cycles each.
http://arxiv.org/pdf/1310.5842.pdf
https://software.intel.com/sites/default/files/article/393199/intel-xeon-phi-coprocessor-vector-microarchitecture.pdf
>>52994945
I think the phi is probably the direction CPUs will be taking moving towards exascale. One CPU for OS and a coprocessor for the rest. Unfortunately, most programmers are shit at parallelism. Have a look at some exascale roadmaps to see the directions people are looking to take this stuff in. Tl;dr current gen stuff is nowhere near good enough.
>>53003519
That other guy is a retard. The Phi coprocessor is nothing like a GPU. Its actually fully fledged x86_64 enabled cores. The code you write for them is completely different.
>>53004013
more to the point, exactly the same as your regular code
>>52994995
how is it less proprietary than radeon?
Can I run SQL on it?
Can you offload Crysis physics onto it?
>>53004127
No, not really. I can see why you'd think that, but your regular code rarely can take advantage of 8 threads, let alone 72. Its incredibly difficult to make efficient code for this thing.
>>52994945
What Linux things can I do with this?
>>52995632
>talking out of your ass
>>52996257
>PowerVR
PowerVR is absolute garbage. Id rather have literally any other GPU
So can I use this if my main CPU is AMD?
do these show up in /proc/cpuinfo?
do these register as CPUs when Linux is starting?
Do they show up in the CPU graphs in htop?
Does Conky have direct access to the CPU usage of these?
Do they do frequency scaling?
>>52997294
>using Windows
>>52997606
what OS? What software is needed to use these?
>>52997606
>icc
is ICC required to make use of these? Can I not use a different compiler?
>>53000017
i bet youre the retard that asked if windows recognized them as cpu cores.
get out
>>53001783
>What's something fun I can do with them that could maybe help me learn something in the process?
encode thousands of movies in a video format that respects your freedom like mkv or ogv and share them on torrent sites with a read me file saying how harmful avi and mpg formats are
>>53004339
I meant more like the regular code you'd write for a multi-core or even multi-node system using OpenMP or MPI rather than the type of code you'd write for, say, a CUDA kernel.
>>53004444
You might be able to cross compile. MPI has support for running parallel code across heterogeneous systems.
I have no idea how this thing works., but it looks cool.
Could this be utilized in a bit mining rig at all?
