|
Here's the page from Intel with the full list of participating vendors: https://software.intel.com/en-us/articles/special-promotion-intel-xeon-phi-coprocessor-31s1p
|
#
?
Oct 29, 2014 23:00
|
|
|
|
| # ? Apr 23, 2024 19:00 |
|
|
atomicthumbs posted:are four 80mm fans really enough to dissipate 1.08 kilowatts of computing Anything's possible as long as you spin them fast enough; the 1U supermicro phi host has ~1KW in 1.75" cooled by 10 40mm 20K RPM fans.
|
|
#
?
Oct 29, 2014 23:33
|
|
|
PCjr sidecar posted:https://software.intel.com/sites/default/files/11MIC42_How_to_Use_MKL_Automatic_Offload_0.pdf Gonna try one of those sometime 
Ika fucked around with this message at 00:12 on Oct 30, 2014 |
|
#
?
Oct 30, 2014 00:04
|
|
|
Ika posted:Gonna try one of those sometime keep in mind that if your matrices are less than 2Kx2K automatic offload won't work.
|
|
#
?
Oct 30, 2014 00:36
|
|
|
I'm tempted to buy one of these just to see how much faster my old MPI jobs will run, compared to how they did on the 4 core Nehalem Xeons I used to use.
Mr Chips fucked around with this message at 01:34 on Oct 30, 2014 |
|
#
?
Oct 30, 2014 01:09
|
|
|
Are there any alternatives for people who want a ton of integer cores?
|
|
#
?
Oct 30, 2014 01:14
|
|
|
PCjr sidecar posted:keep in mind that if your matrices are less than 2Kx2K automatic offload won't work. We run SVD on matrices between 500x500 and 10k x 10k, as well as dense / sparse multiplications with upwards of 100mil nonzero elements. Worth looking into just to speed up processing during debugging / design / evaluation of the math involved.
|
|
#
?
Oct 30, 2014 01:19
|
|
|
No Gravitas posted:Are there any alternatives for people who want a ton of integer cores? Congrats, you're the one person in a million who should be buying an AMD FX-83xx or a similar Opteron. Unless TCO matters more than hardware cost, in which case cram Xeons all up ons.
|
|
#
?
Oct 30, 2014 01:40
|
|
|
If you need to do thousands of integer operations as fast as possible, the Phis or OpenCL devices are still the best options right? I don't really know how expensive Xeon's are, but wouldn't Video Cards be cheaper? Although, is it feasible to make an OpenCL capable FPGA that is integer optimized?
|
|
#
?
Oct 30, 2014 02:47
|
|
|
So I went to a talk by Ray Kurzweil last week, and at one point he briefly dropped that Intel has working prototypes of self-assembling 3D molecular circuits (which he said he has seen in person) and they are planning to have them replace their current production methods before 2020. I dug up some stuff about MIT demonstrating the tech about four years ago, and Kurzweil isn't the type to just straight up lie about something like this, but I can't find anything about Intel specifically. 
|
|
#
?
Oct 30, 2014 05:07
|
|
|
Lord Windy posted:If you need to do thousands of integer operations as fast as possible, the Phis or OpenCL devices are still the best options right? I don't really know how expensive Xeon's are, but wouldn't Video Cards be cheaper? Although, is it feasible to make an OpenCL capable FPGA that is integer optimized?
|
|
#
?
Oct 30, 2014 05:55
|
|
|
Krailor posted:FYI - There's a 3rd party company selling the same model on Amazon for $80. I doubt they ship overseas, but got a link?
|
|
#
?
Oct 30, 2014 06:05
|
|
|
~Coxy posted:I doubt they ship overseas, but got a link? That ship sailed. They did not ship to  while it was up, otherwise I would have gotten one.EDIT: Current best is 140$. They also don't ship to http://www.amazon.com/gp/product/B00OMCB4JI/ref=gno_cart_title_0?ie=UTF8&psc=1&smid=A183W8CNLFPZLY No Gravitas fucked around with this message at 06:12 on Oct 30, 2014 |
|
#
?
Oct 30, 2014 06:10
|
|
|
Rime posted:So I went to a talk by Ray Kurzweil last week, and at one point he briefly dropped that Intel has working prototypes of self-assembling 3D molecular circuits (which he said he has seen in person) and they are planning to have them replace their current production methods before 2020. Ray Kurzweil suffers from a peculiar delusional belief system which leads him to be wildly over-optimistic (kooky, even) about predictions of rapid technological progress. It truly wouldn't surprise me if someone told him something like "We're loving around with this wild idea for making circuits, but holy poo poo the problems are enormous and 2020 is the earliest we could even begin to ship tiny quantities of engineering sample devices to outsiders, if everything goes perfect, which it never does", and Kurzweil heard what he wanted to hear.
|
|
#
?
Oct 30, 2014 07:37
|
|
|
Mr Chips posted:It depends. I'm just curious because I'm learning OpenCL for funsies and I seem to have a 50-50 split between floating point and integer operations. Where do you go to learn more stuff about this? I like benchmarks but I don't want to run my own.
|
|
#
?
Oct 30, 2014 08:46
|
|
|
Speaking of FPGAs, has there been anything new about the Xeons w/FPGAs since that announcement a few months ago?
|
|
#
?
Oct 30, 2014 09:37
|
|
|
BobHoward posted:Ray Kurzweil is crazy http://transcendentman.com/ He takes like a million vitamins a day hoping to stay alive long enough to have his "soul" uploaded to the "matrix" where he will be reunited with his father. No matter that his father didn't live long enough to get uploaded because the future will have computers so powerful they will be able to extract his essence from the fabric of space and time. Cray Cray Super Computer?
|
|
#
?
Oct 30, 2014 16:00
|
|
|
The Amazon Phi deal popped up for a brief while at 92$ shipped. I caved in. Helps that my job will pay for part of it too. I got a forwarding service box and got it sent to there. No idea how (and for how much) I will get it out of the USA, but one step at a time, I guess. Might be something worth investigating for other non-USA people out there. Work is about to get awesome. Once I figure out a way how to exactly cool and feed the thing. Thanks, thread!
|
|
#
?
Oct 31, 2014 00:34
|
|
|
I spent the last couple of days researching the Xeon Phi that is on sale. Man, what a beast. You have those 57 fairly wimpy cores, each running at 1.1GHz. Those cores are very close to Pentium 1. Not even Pentium MMX. Pentium 1. We are talking 1994 here, 20 years ago. Of course they have been tweaked a little bit. We have 64-bit support, some improvements to instruction prefix decode handling, each core has a beefy 512-bit vector unit and they adjusted the pipeline a lot. As a clock speed improvement measure they made it into something like a barrel processor. There are 4 execution threads in every core, for a total of 228. Each clock cycle one of the threads is picked and one or two instructions are issued from that thread. There is a rule that you cannot pick a thread twice in a row, otherwise anything goes. Just because a thread is not picked in a specific clock cycle does not mean it does not make progress. The clock cycle still counts for cache misses, long-latency operations, etc... When some threads are stalled, others can continue. Due to the fact that a thread cannot be picked twice in a row this is like Hyper-threading on steroids. You likely want to have at least two threads per core at all times, if not three or even four. Caches are fun. Each core has the standard 32 kilobytes of data and same amount of instruction cache. 3 cycle latency in practice, sometimes less. There is an L2 cache, 512 kilobytes per core. A better way to think about it is as of a 30 megabyte chunk that is shared between the cores. When your data isn't shared it takes 24 cycles to get data from L2. With shared data that is being modified you can end up with 250 cycles of waiting for a remote part of L2. Don't share non-readonly data. Also only L2 cache prefetches via hardware, so taking some L1 misses is a given. You can try to do software prefetching, of course. Hitting up the main memory? 300+ cycles. Those are the numbers seen in practice. On Intel's paper they are about half and it probably varies between devices. You have 8 gigabytes of GDDR5 RAM at 5GT/s. Want ECC? You got ECC that you can turn off and on... Except it will eat up 1/30th of your RAM and quarter of your bandwidth if you turn it on. Not quite like ECC on the desktop. Which leads me to storage. There isn't any. How do you boot that thing, then? Well, you get a staged bootloader design. Code in ROM boots code in the tiny flash. Flash bootloader signals that the Phi is ready to the host. The host then uploads a Linux distribution to the Phi's RAM and off it all goes. You can store files via NFS or upload them from the host as needed. You can open a serial connection or ssh into your Phi. Something had to go and it is binary compatibility. Sorry, you don't get to run Dune 2 on this. You need to recompile everything in order for it to run. Sure, you can use the bundled qemu if you cannot recompile, but it won't run fast. This recompiling is the first big pain of the Phi. For performance you need to use Intel's C compiler. Sure, there is a GCC port, but it won't use the beefy vector units, won't do the software prefetching and won't be as optimized. This leaves grad students like me a bit in a bind. I cannot run ICC legally as I do get paid for it. However, spending 700$ on the barebones compiler set is a bit rich for me when I earn 3000$ a term. There isn't a free fortran compiler for the Phi at all. There also isn't a simulator, so you cannot test your code once you get it compiled. It will be a fun experiment for me. Everyone loves floating point, but I only require some nice integer performance on the cheap. (Yes, I know I'm doing it wrong, shut up!) The Phi should be roughly equivalent in throughput to 57*2/(7.5*(3.3/1.1)*1.15) = 4 Haswell cores at 3.3GHz, given that it is fed properly. Even for 150$ this is amazing, considering you get the RAM too. And even if it isn't... Well, gotta take fun from something, right? And if running a 57 core, 300W monster won't do it, I don't know what will. Now if only it would get here already, so I could dhrystone, coremark and STREAM it... I will post results here as I get them. No Gravitas fucked around with this message at 21:10 on Nov 1, 2014 |
|
#
?
Nov 1, 2014 20:59
|
|
|
Hello, I'm soon due for building a new heater and I have fallen out of the information loop on what's out there. I have the 2500k in my old computer. I see that with the myriad of tests done it only performs at like 95% as well as the latest 4970k at the highest requirements of gaming. Should I just be transferring my old processor over? The tests I saw use graphics cards that are a little out of my comfortable spending zone (Titans). Is the performance gap wider with more modest graphics cards? (edit: I see the new Assassin's Creed lists the 2500k as the minimum requirement, is that even realistic considering the benchmarks of whateverisoutthererightnow?) I'll be building the computer probably somewhere around March or April of next year. Is there anything I should know about the next processor/product cycle that would make me delay my purchasing spree? And a question with regards to picking a new case: I see Intel has a fancypants new socket LGA 2011v3 or whatever. Did this cause a significant increase in the size of chipsets? Are we expecting a significant increase in the size of chipsets moving into the next 6-7 years? Ernie. fucked around with this message at 21:34 on Nov 1, 2014 |
|
#
?
Nov 1, 2014 21:32
|
|
|
No Gravitas posted:I spent the last couple of days researching the Xeon Phi that is on sale. Man, what a beast. 1) RAM is a bit more than 5 GT/s; ~200 GB/s on the 3xxx, but you have to spread that out across all of the controllers to get anywhere near that. 2) Talk to your local Intel sales rep and see if he can help you with a license; see if he can get you a VTune license also. If you have a local/regional HPC center that has access to recent Intel compilers they may have the cross-compilation tools; you may be able to compile there and copy binaries to your desktop. If you're in the US, try to get on XSEDE; your campus champion can get you a startup allocation fairly easily, which will get you access to TACC's Stampede cluster (and Phi tools.) 3) If you haven't seen it yet, https://software.intel.com/en-us/articles/intel-manycore-platform-software-stack-mpss 4) The STREAM benchmark results site has specific tuning/compilation examples for the Phi.
|
|
#
?
Nov 1, 2014 22:43
|
|
|
Ernie. posted:Hello, I'm soon due for building a new heater and I have fallen out of the information loop on what's out there. Your 2500K is almost as fast as haswell and can keep up just fine if you overclock. If you need a bit of oomph you could get a new video card. Check the PC Building & Parts Picking megathread for more info about specifics.
|
|
#
?
Nov 1, 2014 23:20
|
|
|
Rexxed posted:Your 2500K is almost as fast as haswell and can keep up just fine if you overclock. If you need a bit of oomph you could get a new video card. Check the PC Building & Parts Picking megathread for more info about specifics. Sorry for my misposting. I thought since all of my questions were CPU-related I'd just post them here. Thanks for the help, regardless! 
|
|
#
?
Nov 1, 2014 23:31
|
|
|
No Gravitas posted:phi stuff The only annoying thing is that most boards don't (explicitly) support it unless you already are running something like an E5.
|
|
#
?
Nov 2, 2014 00:33
|
|
|
PCjr sidecar posted:1) RAM is a bit more than 5 GT/s; ~200 GB/s on the 3xxx, but you have to spread that out across all of the controllers to get anywhere near that. 2) Nah. I don't like to involve Team Blue any more than I need to. What I'm planning to do is a terrible abuse of the Phi's capabilities. I'm basically Aperture Science in real life. I can't help it, I just like trying crazy things and making them work. I can imagine how the conversation with Intel would go.  Hi, I'm a poor grad student doing work at a university. Can I have a compiler for your 2000$ device that was on sale? I will earn some money with it by doing integer-intensive computations totally unsuited to the Phi! Please talk to your datacenter manager and ask him to buy you a licence. I don't have a datacenter. I'm running it all by myself. It's a blast. How the hell are you running it without a datacenter? In my desktop, I have a sever motherboard. How in the world are you cooling it? Oh, I have a quiet fan, duct tape and some ductwork made from card stock.  Can we hire you? We need enthusiastic/excited madmen like you who push the envelope of sanity in name of being cheap. We can put you on our circuit testing. If you cannot break something, nothing will.Yeah, I think I will just run with the GPL stuff they released. Which reminds me: No gprof either, but you do get gdb. 3) Yup. Already using it. I have hello world ready and waiting to run. The Phi will get here in a couple of weeks at the earliest. For now I'm working on cutting down Octave to build without fortran for now... Loooong story. Ika posted:The only annoying thing is that most boards don't (explicitly) support it unless you already are running something like an E5. I'm running on a Xeon E3-1226 v3, and a server motherboard that has all the right BIOS thingies to support it. Fingers crossed! If it does not work, I can always sell it or keep as a cool doorstop.
|
|
#
?
Nov 2, 2014 00:39
|
|
|
No Gravitas posted:I'm running on a Xeon E3-1226 v3, and a server motherboard that has all the right BIOS thingies to support it. Fingers crossed! If it does not work, I can always sell it or keep as a cool doorstop. Good luck, I'm still trying to figure out whether I can upgrade my dev box to something that supports the phi while remaining in an acceptable budget, since the performance increase is an unknown.
|
|
#
?
Nov 2, 2014 00:54
|
|
|
I suppose there's some Hadoop native code available for the Phi from Intel then? Most people don't want to switch Hadoop distros though either and would just rather have some native code that's got Java bindings, but Knights Landing being aimed at the research rather than commercial crowd hasn't gotten there in popularity. Also was thinking of adding Storm stuff if nobody got to it. Seems pretty well suited for dumb jobs that need tons of memory bandwidth that you can't get normally on a general CPU without needing to go down to CUDA and deal with Tesla GPUs. Last I saw I saw no point looking for Tesla crap for my purposes because consumer GPUs were still better price / performance. But my interest is piqued....
|
|
#
?
Nov 2, 2014 12:12
|
|
|
No Gravitas posted:2) Nah. I don't like to involve Team Blue any more than I need to. What I'm planning to do is a terrible abuse of the Phi's capabilities. I'm basically Aperture Science in real life. I can't help it, I just like trying crazy things and making them work. I can imagine how the conversation with Intel would go. Sure, who cares, take 5, under 100k/month volumes isn't worth my time to fight
|
|
#
?
Nov 2, 2014 17:10
|
|
|
If the Intel rep is anything like our Cisco rep at work they'll throw licenses at you. Dude I work with was fretting about a couple $200 annual licenses for his ASA (firewall) and the rep was like "Oh it's not a production device? How many unlimited licenses do you need?"
|
|
#
?
Nov 2, 2014 18:33
|
|
|
No Gravitas posted:I can imagine how the conversation with Intel would go. Seriously, they want you to use their poo poo so when you go out in the real world you'll (hopefully) evangelize for it and they'll get some big sales. A guy at my old lab got a Quadro K5000 with like a half page proposal to nvidia; the only proviso is that he has to acknowledge them in any papers he writes.
|
|
#
?
Nov 2, 2014 19:15
|
|
|
The old drug dealer model modernized - "development at scale is free!"
|
|
#
?
Nov 2, 2014 21:39
|
|
|
Heh. I wish it were only true that they'd send me hardware too. I mean, Knight's Landing is coming out soon, right? Atom is surely better isoclock than Pentium, right? Maybe it even takes unmodified binaries too? Please? As much as I'd love to get stuff, I won't be asking for it. 1) I will earn money on this. 2) I'm abusing the Phi. 3) It isn't even needed to use the Phi. 4) I'm not even gaining that much by using the Phi, if my estimates hold. I'm doubling my performance at most. Aaaaand... It shipped. 2+ kg package. Lovely. I think the heatsink alone must be 1kg. Assuming I manage to get it out of the USA somehow it will be here by late November. I cannot imagine how much the forwarding of that lovely brick will be. By the time I get it I hope to have Octave cut down to make it all work. What have I done. This is insane.
|
|
#
?
Nov 2, 2014 21:48
|
|
|
No Gravitas posted:For now I'm working on cutting down Octave to build without fortran for now... Loooong story.
|
|
#
?
Nov 2, 2014 21:56
|
|
|
No Gravitas posted:1) I will earn money on this. Yeah, Intel's really concerned you'll take a chunk out of their bottom line with a single 2-year old chip on 3-year old process tech. edit You're making up a world that doesn't exist and crippling your project because of it. JawnV6 fucked around with this message at 23:16 on Nov 2, 2014 |
|
#
?
Nov 2, 2014 21:58
|
|
|
Professor Science posted:you have piqued my interest, why are you doing this My boss is too cheap to use MATLAB, has me using Octave instead. We don't really need MATLAB/Octave. No math functions are actually used. Pure scripting. poo poo runs too slowly, no surprise. The code is effectively written as if it were C, just happens to be in Octave. This will never run fast. We have 2-3 kloc of this. I wrote a fair chunk of it too. It is the best documented code I have ever worked with in my life. It is beautiful. All variables have good names. There are regression tests. Power-on self-tests even. Abstractions are great. Consistent style... and beyond horrid choice of programming language for the task. As you can see, the project is quite absurd. I gave in and embraced the insanity. I thought that since the Phi is on sale, maybe I should just play along and propose something absurd right back. I love hardware and they won't hear about anything other than Octave. I proposed a 50-50 split on the costs of the Phi to (maybe / at most) double the total processing speed. I get to keep the Phi once we are done. I pay for the Phi, they pay for the electricity and the power supply. To my horror I got a yes. Nevermind that for that money you could probably buy MATLAB, but well... I don't even care anymore. So cold. There is no free fortran compiler and since we are only scripting... I'm slowly filing away all the fortran bits that will never get used... EDIT: I forgot the best part. The abuse of the Phi that I keep talking about? Not running a single application with 100 threads, oh no. I will be running ~100 single-threaded instances of Octave on the Phi. The boss is always right. The boss is always right. The boss is always right. Why does it hurt so much that the boss is always right. I begged and pleaded. I asked for MATLAB. For C. For anything. The boss is always right. I just mention that we could buy the Phi for a tiny boost and she just home ran with it to death. JawnV6 posted:You're making up a world that doesn't exist and crippling your project because of it. You know what... You have a point. I will try. Can't hurt too much. No Gravitas fucked around with this message at 23:07 on Nov 2, 2014 |
|
#
?
Nov 2, 2014 22:56
|
|
|
Matlab is something like 1k or 2k for the base suite, and if you want to do anything other than multiply vectors / matrices and maybe basic functions on matrices you need to buy the correct addin toolbox for another 500.
|
|
#
?
Nov 2, 2014 23:45
|
|
|
Ika posted:Matlab is something like 1k or 2k for the base suite, and if you want to do anything other than multiply vectors / matrices and maybe basic functions on matrices you need to buy the correct addin toolbox for another 500. Well, throwing the 100$ Phi at the problem makes a bit more sense then. Still not as much sense as just porting over to Python, C, Julia, Java, etc...
|
|
#
?
Nov 3, 2014 00:11
|
|
|
No Gravitas posted:EDIT: I forgot the best part. The abuse of the Phi that I keep talking about? Not running a single application with 100 threads, oh no. I will be running ~100 single-threaded instances of Octave on the Phi. my god, what have i unleashed
|
|
#
?
Nov 3, 2014 00:58
|
|
|
atomicthumbs posted:my god, what have i unleashed
|
|
#
?
Nov 3, 2014 01:32
|
|
|
|
| # ? Apr 23, 2024 19:00 |
|
|
necrobobsledder posted:I suppose there's some Hadoop native code available for the Phi from Intel then? Most people don't want to switch Hadoop distros though either and would just rather have some native code that's got Java bindings, but Knights Landing being aimed at the research rather than commercial crowd hasn't gotten there in popularity. Also was thinking of adding Storm stuff if nobody got to it. Seems pretty well suited for dumb jobs that need tons of memory bandwidth that you can't get normally on a general CPU without needing to go down to CUDA and deal with Tesla GPUs. Last I saw I saw no point looking for Tesla crap for my purposes because consumer GPUs were still better price / performance. Moving data on/off card efficiently is difficult and the non-vectorized/non-optimized performance is really bad, like "smoked by a single $400 E5-2620 at 1/5 the power" bad. At $150 it's a good development platform, and good experience for thinking about running on KL in 2016-ish, which will have much better performance for that kind of workload.
|
|
#
?
Nov 3, 2014 15:50
|
|