Viewer-Mail Ep3: VGA Shaped Intel CPU with 61 Cores - Intel Xeon Phi

Support me on Patreon:
/ der8auer
---------------------------------------------------------
Save 10% on your iFixit purchase: DER8AUER10
iFixit: ifixit.com/Der8...
Buy the Pro Tech Toolkit: bit.ly/2JOFD8f
Buy the Manta Precision Bit Set: bit.ly/2JSpkYa
---------------------------------------------------------
Find my products at Caseking:
Delid Die Mate 2: bit.ly/2Rhv4y7
Delid Die Mate X: bit.ly/2EYLwwG
Skylake-X Direct Die Frame: bit.ly/2GW6yyC
9th Gen OC Frame: bit.ly/2UVSubi
Debug-LED: bit.ly/2QVUEt0
Pretested CPUs: bit.ly/2Aqhf6y
---------------------------------------------------------
My Equipment:
USB- Microscope*: amzn.to/2Vi4dky
My Camera*: amzn.to/2BN4h2O
(*Affiliate Links: If you buy something, I will get a part of the profit)
---------------------------------------------------------
Music / Credits:
Outro:
Vendetta Beats - Angelus
(Custom Version)

[@JN.0_o]

I'd like to take a moment to convey my appreciation and admiration for how you make 2 separate versions of each of your videos in English and German, and somehow manage to keep them so consistent? That seems like it would be extremely difficult to do

[@sirfairplay9153]

i second that, the guy is a legend!

[@guiorgy]

Personally, if it was me, I would have recorded it once, and replaced the voice in post for the other language. I still think they should do so in cases where they are limited in recording time, like *CES* and such. If you can only record for say 10 minutes and you spend 5 for each language, all viewers suffer with rushed and shorter content. Doing the second language in post is a good compromise imo

[@mscd9676]

I had to make my resume in both English and Polish and keeping them consistent wasn't that easy

[@spacecy]

he is awesome!

[@TheVeratrix]

He's German that's why.

[@vishwajeetjangra3390]

i just realized that you record the same video two times for English and German That's some real hardwork and consistency

[@samiulislamsharan]

he's a really patient person you have to admit that.

[@TheOz1999]

@Joseph Richter yeaaah, wouldn't it make more sense to still upload the two videos in different languages but to two separate channels instead, with each channel corresponding to the language?

[@excitedbox5705]

@Joseph Richter no due to duplicate content filters. It would be better if youtube allowed putting different audio tracks on 1 video. DVDs have had that feature for over 20 years.

[@TheGoldgoat]

German quality

[@UmVtCg]

You just realized that, ahw... that's cute.

[@ethan16384]

I know I'm a bit late, but GPUPI can actually run on any Xeon Phi! You need specifically version 3.3, after that they stopped supporting Xeon Phis. If you'd like, I can get you in contact with a guy who holds a few of the hardwarebot GPUPI records with Xeon Phis

[@mikebutler9332]

I do scientific computing. I've never worked with them myself but I see these floating around on ebay at reasonable prices and was tempted to try and adapt them for workstation usage. It was a pleasant surprise to see this show up in my feed. I know there is an openCL SDK floating around for these. I've heard mixed opinions about it but you could compile something like blender with their openCL implementation to compare performance. May or may not be a headache though. It can be a drop in replacement in some instances.

[@TheBackyardChemist]

I think the opencl sdk either never left beta or was just really buggy and clunky. Either way, Intel cancelled its development soon after release and instructed everyone to just use the Intel compilers.

[@justacomment1657]

@@TheBackyardChemist Intel Basicly cancelled the whole thing.... Just look at the latest driver release for those... They are massively crappy as they are not add in CPUs, rather separate pcs with next to no i/o ability... Feeding them data and getting it back is pure horror

[@christophertstone]

Phi's suck at OpenCL-type loads, get a GPU if that's your application. Phi's excel at applications with tons of branching threads. Most scientific computing isn't this.

[@mikebutler9332]

@@christophertstone I didn't say they were. I implied the opposite in fact. I was mentioning a possible workload that someone could use for testing. If nothing else it might be an accessible way of testing thermals.

[@carmonben]

The reason alot of people refer to these as gpus is because they started off life as gpus If you look carefully you can still see the empty port where an HDMI would have been. Second for benchmarking them, the only benchmark I was able to run on mine was the intel linpack benchmark, and I stopped it before it finished because it was shaking my table from how fast the fan was spinning (7120A)

[@yogi.g]

Intel: This GPU is a CPU. Nvidia: .......

[@NonsensicalSpudz]

Nvidia - Always has been

[@MoraFermi]

Intel: This CPU was supposed to be a GPU but we never could get the renderer code to work.

[@ayuchanayuko]

Larrabee intensifies

[@betadoctor]

@@NonsensicalSpudz yep.

[@Jaker788]

@@MoraFermi I think bit was more a performance and power thing. You can make a driver that gets 3D rendering to go the the AVX-512 instructions. It's just wasn't as good as Intel though it would be as a GPU.

[@chubbysumo2230]

So I believe the reason that Intel chose the thermal pad on the coprocessor instead of thermal paste, was one of long-term usage. Expected these to be used in a hot server room for a very very long time, and they did not want the thermal performance of the cooler and heatsink to degrade over time, causing there big corporate money Spenders to be unhappy. These things were not cheap when they first came out.

[@JmbFountain]

If your server Room is hot, you are doing something wrong. It should be around 17°C to 19°C

[@ryanbernard6550]

@@JmbFountain i know you're right but god dam I've never been in a cold server room

[@myownsite]

@@JmbFountain many data centers run warm because its efficient

[@JasonB808]

Don’t you mean server farm. Companies that would have these co processors would have Server Farms. They are now called Data Centers or AI/ML Centers. They would be definitely cooled with high performance Air conditioning with humidity control. Though it is true that Thermal Pads will last a lot longer than Thermal paste. We can’t think like a Gamer. These are for Pro users.

[@patricklegault6383]

@@JmbFountain actually. server room are starting to adapt to warmer temperature as it saves them money. here, in my area, they run from 20 to 22c the hardware works just fine they have hot and cold rows. the cold one is theses days around 20C while the hot one is rather hot... can range from 24 to 30C easily

[@jetfu400]

2013: 60+ core cpu exist 2020: heres our all new 16 core cpu at a price that youll die for.

[@zacht.1531]

To be fair those 60 CPU cores are really really really slow and shitty

[@yujenlin1197]

2020 Amd: here's a 64 core cpu

[@shadowguardian3612]

750$ for 16 cores isnt bad

[@christophertstone]

@@zacht.1531 They aren't so much "slow and shitty" as they are _simple and application specific_ Phi's have 57-61 cores, with 4 threads per core, running at 1.0-1.2GHz. Per thread they're 1/10th the performance of a modern AMD64; but there's up to 244 threads running simultaneously. They're best suited for applications with massive numbers of threads that benefit from packed operations and per-thread branching, especially where the branching may cause pipeline stalls that kill performance in low-thread count CPUs. The other thing these are not is an OpenCL/CUDA replacement. GPUs are really good at deterministic programming (ie, no branching and no easy way to work around no branching). If you're doing most "math" (like most scientific applications), GPUs will be much faster than Phi.

[@alexisleftist778]

i mean you can buy a 64 core threadripper too now , its like 4000 dollars but its damn near as fast as any of the other cpus except it has more cores. kinda amazing t hat they got a 64 core cpu to be just as good at gaming as their lower core count cpus

[@lost4468yt]

There's plenty of tutorials out there on how to boot Linux on the coprocessors. Intel MPSS supports SUSE and RedHat as standard. And you can get Ubuntu running quite easily. Once up you can setup SSH on it and just SSH in like a normal distro.

[@Mythricia1988]

Hmm, I thought these Phi's were usually used via remote access from the host computer. Basically you remote-desktop (or SSH into them, really) into the card, from the host. At least that's how I have seen them used. I think Intel MPSS is the tool used to set up the card and create this local PCIe network to the card, and is how you control & monitor it. I don't have any direct experience with it though myself!

[@Vatharian]

It depends on firmware. You can boot them into slave mode, but it definitely won't work under Windows.

[@andreaslegomovies]

@@Vatharian I'm ssh'ing into and offloading onto mine from a Windows 10 host (linux requires too old kernels). You don't have a compiler in the tiny linux env on the card, so you have to cross-compile on the host and send it over the pcie network to the card. (Some Intel tools can do this efficiently bts through openmp and similar interfaces)

[@madcatattack1]

This is wonderful. I just love it when slightly older technology is still very relevant for someone's use case.

[@brentboswell1294]

Intel also made servers (Knights Bridge and Knights Mill) that used these processors. They were "compute sleds" designed to go into a Bobcat Peak chassis (2U chassis designed to hold four sleds). They were intended for supercomputing type applications, where the number of threads was more advantageous than raw processor speed. The instruction set is x86, you can install a standard Linux build on it!

[@uzefulvideos3440]

Those Phi coprocessors also run their own small Linux operating system, which you can log into over SSH. You don't need special code then, most x86 Linux code should work.

[@coopercoldwell]

A really cool proof of concept would be to get a Minecraft server (or something similar) running on one of these.

[@TheDestoyer24]

Shiek was like- PAY ME ATTENTION; I KNOW MY LEG ISN'T AS BAD BUT I LIKED THE LOVE I GOT WHEN IT WAS WORSE- PET MEEEEEEEEEEE

[@der8auer]

🤣🤣

[@twistedridermike]

I've looked into these a little, they seem to run a CentOS kernel internally. I think you SSH to them from the host system. I'm not sure if software is copied to the device normally however. I'd like to know if these could be added as an MPI based cluster in inside the host box. With two of these in a knights landing host you'd have a massive amount of x86 cores.

[@johnny1968a]

This was a very interesting video. I really like the explanations. So interesting to see such a big cpu on a card!

[@charleshines6155]

Is there a similar card that uses AMD CPus with high core counts in a similar way? If so it would be interesting to see that. Imagine something similar with the latest EPYC processor on it.

[@2001pulsar]

Glad to see sheik is getting better. Poor pussey had too much adventure and lost one life.

[@tacticalcenter8658]

They use the thick thermal pad on a vapor chamber because it's not a very good seal using thermal paste alone. Some times different heat sinks requires different types of materials. And even then some require different consistencies of that type of interface materials. You can't always use thermal Grizzly for every thermal interface material in every situation. In any case, if I remember correctly the vapor chamber can't withstand huge mounting pressures that are idea when using paste, so the heat pad is more idea in this situation.

[@Psyduck1989]

Im very interested with the temp, u got so much radiators installed.

[@whiteburr]

Agreed. I have more rads area than this on ryzen and single 2080ti, I can barely increase my coolant temperature. But it would be interesting to see a follow up on this setup

[@basedanon420]

@@whiteburr on 1 x 420 and 2 x 360 rads (both thick) I have difficulty reaching 5 degrees delta T even with insane amounts of stress on my system (furmark + small ffts avx2). This is on a direct die 9900k pulling 250W and 2080ti pulling 400W+

[@kleanthisgroutides7100]

Insane engineering on these, imagine XE if they nail it.

[@ChristianStout]

Where'd you find those waterblocks? I've been meaning to get a cheap Phi on eBay to mess about with in my X99 system, but I can't find a good thermal solution for the older ones.

[@hjq0023]

Can this coprescessor work with a gaming X99 mobo? I already have a Xeon processor installed in my X99 mobo

[@stageselectca]

would have been nice to see before/after thermal performance, but holy crap that was a shock seeing how big the case was all of a sudden haha

[@gsuberland]

The Phi shows up on the host as a number of different devices exposed over PCI-e, including a network card so you can do "remote" login to it over SSH. You can manage a Phi from the host using Intel's Manycore Platform Software Stack (MPSS) tooling on Windows or Linux. The processors are MIC architecture, which is x86-compatible but won't run the same generic x86 code you'd build for a modern desktop CPU because the instruction set is heavily reduced. Puget Systems have a few good blog posts on how to set up and use a Phi under Windows. If you want to write code for it, you really want Intel's compiler (Parallel Studio XE), but it's a commercial product. You can get free copies of it if you go through their qualification program (students, educators, FOSS projects, etc.) but for the average joe it's at least $700 for a license. You can also build code for it with gcc, targeting OpenMP, but that's much less optimal than Intel's option and I'm not sure how well that part of gcc is maintained these days. Once you've got binaries to run on it, you just SCP them over, or use one of the MPSS host commands to copy the file.

[@Jail_Brreaker]

Aha ich glaub da wurden die video Titel vertauscht :D

[@matthis1475]

Gleicher Gedanke Mal gucken wann er es merkt

[@Omega-oo4hh]

Bester Mann dachte auch schon aber kann man sich auch so geben

[@jondonnelly4831]

I love the tubing and fittings. Looks Great.

[@markmanning2921]

Churck Moore wrote a VLSI editor in color-forth then used that to develop 144 core CPUs which are microcoded in forth. intel is a multi billion dollar a year mega corporation and chuck moore is one man.

[@P4GrAnGeR]

Easiest way to use them is to upload a "normal" linux distro into them and run the benchmarks directly there. They're just PCs, it's just annoying to talk to them.

[@justacomment1657]

The annoying to talk to part is the reason they are dead. Scaling is horrible, overhead massive...

[@dabnation2534]

love this series please keep it up. these are very interesting and fun to watch.

[@TheBackyardChemist]

I am glad that there are research groups that can make use of these cards. We certainly cannot. I have access to 2 supercomputers that were (for some godforsaken reason) built with a bunch of these accelerator cards. While I cannot speak for every user, I have a strong suspicion that noone actually uses the Phi cards, because software support is nonexistent. These things cannot run OpenCL, because Intel cancelled the development of the OpenCL SDK shortly after release, and of course they cannot run CUDA. The only way to get any code running on these is to recompile your software from source code, using an obsolete version of Intel's compilers, as support was completely removed from newer versions. Even if you can get something to run on the cards, they are about as fast as a glacier before climate change, and you have to rewrite everything from scratch to get reasonable performance. Its months of work, very few people can do it, and many scientific codes are a poor fit for the Xeon Phi architecture, so there is no guarantee it will be ever worth the effort to rewrite the software. If they just spent all that money on more CPU cores, they would be used a lot more, instead they wasted probably about a million euros of taxpayer money on these stupid things that almost noone here can use.

[@Lady_Zenith]

You are right, they are useless and pretty much always were useless. At best scenario you are looking at 61 old Atom-level cores that are really weak at everything except specific AVX code, but that same code can be written to run in OpenCL/Cuda much faster. This thing got trashed by GPUs even when it was new, now its complete goner. And as a CPU this whole thing is a joke. If it threads ideally you can see it go up to like somewhere between 8C-12C ZEN2. Any hiccups in thread scaling = too slow to even bother. Not to mention the very limiting factor that is the memory size. Intel rightfully got nowhere with this thing, but they still had so much leverage that a lot of them got pushed in to various systems where they made questionable sense even when brand new. Tryning to run those things now is like trying to optimize for CELL or Itanium or Netburst. Just no. Its slow, old, let it be.

[@TheBackyardChemist]

@@Lady_Zenith Yeah, I am just so pissed off that my government choose to spend the very very limited national HPC budget on these stupid things, instead of more ivybridge and haswell cpu cores.

[@Lady_Zenith]

@@TheBackyardChemist To put it in to perspective, the top latest model of this does around 1600-1700 points in cinebench R15 when the OS itself is running on it. Decent for 2013/14, but the top-end Haswell-EP from also 2014, the 18-core 2699v3 can do around 2600-2800 at lower TDP, and this is ideally threaded scenario, at pretty much anything else the Haswell will demolish it.

[@sunyamacs]

Roman: touches heatsink with bare finger and leaves fingerprints on it Me: NOOOOOOOOOOOO

[@samfedorka5629]

5:52 that is one oddball VRM. We have 4 phases on the top left and 4 more on the top right that I assume are all vcore. (the IR VRM controller at the top appears to be an 8 phase) The 4 on the left share a power input, but the 4 on the right are split between two power inputs. The we have one on the left and 2 on the right for the power plane on the top of the PCB, which I assume is for memory. The one on the left has the same power input as the vcore above it, but of the two on the right, the upper one shares an input with vcore, then there's a gap with a phase not related to vmem, then the bottom most phase is another memory one, but it has its own input that's seperate from the rest. Finally the lowest phase on the left (EVID2) and the second lowest one on the right (EU7D2) are not related to the memory and are likely not related to vcore either. Minor rails? There is a third smaller VRM controller that looks like it might be powering those two as two seperate one-phase VRMs. Regarding the power inputs, you can see two phases for VCore, one for mem, the one unrelated one all connect to the lowest 6-pin. You can also see an inductor cleaning up the noise near each of the two 6-pins and there's one on the far left. Maybe there's some noise added to the power plane as it goes across the board? Maybe everything there is running off PCIE power (unlikely). Then all the capacitors are on the back, but almost all of them are for vcore.

[@JMUDoc]

I love the way Roman pronounces "k-zeon" :)

[@EspHack]

I mean.. is it even a Cpu? I guess the C can be kept since its a "co-processor" rather than "central"

[@laurelsporter4569]

It can run Linux, and act like an independent computer. It's a fully fledged SBC.

[@bryantallen703]

Talk to Wendell over at "Level1Techs". If anyone would know how to benchmark or utilize Phi, it would be him. He might surprise us all and get'em running in Linux. He had some IBM G9's running and benchmarking them.

[@nutzeeer]

ich hab mal ne frage: wie sieht es eigentlich mit pin kompatibilitäten bei GPUs aus? hat jede version ihre eigene pin-belegung, oder ist das in jeder generation einheitlich? könnte man ausprobieren eine kleinere gpu auf ein großes board für eine bessere energieversorgung zu geben, um das overclockingpotential auszuloten? wie gut funktionieren große chips mit weniger versorgung? und mich würde noch interessieren welche performance ein kühler der gtx 8800 heutzutage bringt

[@akhlaken]

You can find benchmark ( flops ) in linux quite easily for these Phi Coprocessor using Intel MPI, OpenMP and Intel MKL. Benchmark can be run in either native where the benchmark run/execute on Phi itself ( these Phi run very small linux ) or in MIC-Offload where it act as accelerator where the dataset can be larger. They have either 8GB or 16GB ECC GDDR5 I don't remember these Phi having graphite pad, hundreds that I've worked with comes with thermal paste. Maybe the sender have tried to improve the temperature by using the graphite pad.

[@andreaslegomovies]

Mine also had three graphite pad, 7120P

[@emboop2346]

I've always wondered what those cards looked like torn down thanks for such a good video

[@chrisalarcon6982]

Do you know which motherboards use that card? Looking at building a super computer, but see that these type of cards are limited and based on BAR size along with type of mother board.

[@robinliu5849]

Could you please share a bit more on the cooling solution, especially where did you get the cooling case? I am trying to fit three of them into the tided slots but the blue large Intel case is annoying...

[@thomascooley2749]

Glad shiek is doing better hope to see more of her soon

[@Karthig1987]

Awesome that shiek is doing much better now. Was quite a sad story in the last video. Xeon phi seems like a really cool product for the enterprise space.

[@blazetechstuff]

Doesn't ecc memory need an additional chip/register?

[@blkspade23]

They should function as OpenCL devices. Something like Hashcat might be a fun experiment to run on this.

[@Arctic_silverstreak]

So what does this actually for? I found one not similar, little bit on the lower side for around $60 Just for fun and experiment

[@kaloy066]

Shiek: Meow 😽 good thing he's been recovering

[@GeoffSeeley]

@7:38 Thank you for the Shiek update!

[@matonted]

You could also technically use windriver which should be able to generate libraries for direct acces to all the HW and write your own program with that. If you have access to intel's documentation that is. www.jungo.com/st/products/windriver/

[@hayzeproductions7093]

Havent done anything with these before, and prices have came down a ton, but theoretically you may be able to use these with virtual machines and run plenty of VM's on them. Splitting the core count up.

[@shaggy8736]

Hab da mal ne frage , bau ein neues amd system mit dem ryzen 5 3600 , und 32 gb Ram auf einem Matx, reicht der Boxed Kühler auf dauer aus auch oder ist man gezwungen ein neuen lüfter zu kaufen , da die boxed lüfter ja nicht den besten Ruf haben bezüglich der Kühl Leistung

[@dorion9111]

Those cards can be utilized really easily under Linux

[@der8auer]

I remember last time I tried installing the Nvidia driver for RTX2080 Ti in Linux it took me 4 hours

[@SHUJINCELL]

Stop snitchin! 😅😅

[@Paginski]

@@der8auer which distro? On something like Manjaro it's super easy

[@jorno1994]

@@der8auer on Pop!OS, Mint and Manjaro it should be as easy as clicking a check box.

[@monchiabbad]

@@der8auer recently under ubuntu(-mate) 20.04 the "aditional Drivers" tab in the "Software & Updates"-settings tab after some waiting should detect most less ancient graphics cards from NVidia and give many driver options. And on many laptops intel igd/NVidia powersavings switching should work pretty much upon reboot after installing from the "Aditional Drivers".Installing from the driver settings window should work perfectly now. Even though of course it is not fully opensource ( there is still the BLOB ) it should work as safe as the BLOB under windows.

[@custume]

can't you use that CPU on virtual machines (like a dedicated azure server) ??

[@Jaker788]

I don't know. But I know the cores are really weak for most applications. It's a just an AVX-512 machine for the most part

[@Tyrian2k]

I think Shiek was trying to say "I'm OpenMP pro! give me keyboard!"

[@no-eb2xx]

Cooler idea: Place the RAM and VRM on the back and use a bent vapour chamber to tranfer the heat to the other side of the PCB for active cooling, the actual GPU itself could be really short since you put a lot of stuff on the back. or make holes into the PCB and transfer heat trough with a few heatpipes connected to the active cooling. Could that be done?

[@popcorny007]

Just having the chips on the backside of the PCB is almost too tall already. Any backplate/heatsync on top of that would have to be extremely thin to not be too tall.

[@no-eb2xx]

@@popcorny007 ehh screw the standards, they can make it fit no problem. Or well just mount the gpu into the second slot, problem solved.

[@jamesi6737]

Why do you use sequential cooling fluid flow? This is definitely less effective due to temperature grow from one unit to another.

[@James_Moton]

Why does this specific CPU model have only 61 cores instead of a full 64? What happened to the other 3 cores?

[@patrickprafke4894]

Not having something like this available to consumer use is something us home lab people could use. We missed out.

[@FARBerserker]

Would it not be better to have dedicated cooling cycles for CPUs and "GPUs"?

[@raphaelsander4575]

Hi, can I asign some cores to virtual machine (VMWare or VirtualBox)? Could you make a video showing it?

[@9iyushy838]

I think you uploaded wrong videos under wrong title

[@corason4444]

9:19 AAAAAAAAAAAAAAAAAAAAAAAA WFT IS THIS BEAUTY??????!!!!!!!!11111111

[@williamhart4896]

Dang for parallel processing rigs that xenon phi co processor is a brute force tool with the right programming language .

[@isilder]

The metal connection from cpu side to the backplate side is not going to help cooling much..its too far ..theres already heaps of contact with air before that point .. water cooling is a bit extreme .. how about just put holes in the blue cover to let air flow in ... eg Attach a fan over a hole in the blue cover .. or. To the end of the card .. or just sure the air pressure inside the case will cause air flow through the blue cover to the outside ..

[@andreaslegomovies]

I've had the 7120 xeon phi in my personal system for about 3 or 4 years. I'm very proficient in C++ and OpenCL, but I haven't yet found a workload that runs efficiently on these. Modern GPUs (GTX 1070 and Vega 64) and CPUs (6800K and recently 3960X) has been much faster at everything I've thrown at them. The best relative performers has been very branched OpenCL code, but the OpenCL drivers are old (2014 is the latest update) and not optimised well, so the performance is still ass.

[@axe863]

I would assume they're better with random forests than gpus.

[@jonasvonsolem]

ANSYS 16 should be able to use the phi's to accelerate CFD and/or FEA simulations, remember seeing a brochure about that.

[@plower221]

First - that is an awesome looking build / cooling loop. Do you have a way to measure the thermals of the package without writing custom code? Also, what happened to the kitty??

[@canfarts784]

can those be used to improve your gaming experience or perhaps something like video encoding?

[@EmreA]

What about eCPU setup via thunderbolt?

[@zsoltpeterdaniel8413]

Same as u would with gpu but only pci e x4 so idk what would it do with performance

[@EnvAdam]

5:19 - I alwasys thought this was interesting with the cavium thunder CPUs because they are 32 cores but 128 threads.

[@moebius2k103]

Great video, very cool to see some uncommon hardware. One question though, what is that tubing and what are those fittings that you used?

[@fabienleydier]

I don't know about GCC compiler for those, but with Intel compiler (not free...), you have to target the device with the option -mmic nativeMIC. The code should run as if it was a manycore system. To run on both the CPU and coprocessors, you have to code the actual offloadings by hand! So you can compile chemistry softwares like NAMD or GROMACS, they support it already and are under free licenses of some sort (GROMACS is much easier though imo). During the run, you can ask for the device usage on Linux, but quite frankly, I don't remember how to do it (you can't just "htop" on those!).

[@lastfirst5863]

What exactly is the use case for coprocessors like that? Could you make a VM run on one? Could you hook it up to discrete graphics?

[@caneslotocato3382]

It's not possible to pass the phi as processor with a valid hypervisor?(proxmox/vsphere)

[@NinerFourWhiskey]

I'm a bit familiar with the Phi copro cards. They're meant only for server cooling. Rather than all this liquid trouble, it would be easier to acquire a SuperMicro 4U pedestal case and their kit for passive-cool cards, like Phi or Tesla. Or even a bare-bones 4U server chassis. These are fairly plentiful on the used market and cheap, as are Xeon E5-v4 boards and cpus. From a software perspective, unlike GPU, phi are cut-down x86 and there's a special compiler for them. The simplified idea is to write regular x86 code for your compute tasks, then compile it through the Phi compiler and scale it out amongst as many cores as you can afford to operate. Honestly, I found the NV GPU easier to write code for, much more stable, with the Volta having better fp64 and the Turing having better fp16.

[@sinephase]

wait, so you think two* 360 radiator and loop will handle over 600 watts?

[@TheFakler]

There is a an open source scientific program called NWchem, which has benchmarks you can run on the xeon phi.

[@deltacx1059]

Someone needs to get crap like this working for gaming, it could be so useful for calculating physics.

[@mickolesmana5899]

Did you mean physX?

[@deltacx1059]

@@mickolesmana5899 yes and no, it would need to be a bit more robust.

[@mickolesmana5899]

@@deltacx1059 isn't already robust? but developer (in my assumption) don't utilize physX function as much

[@deltacx1059]

@@mickolesmana5899 it would just be nice to have something that works more naturally with the newer APIs and engines

[@mickolesmana5899]

@@deltacx1059 Ah you meant is ASIIC Card for PhysX that being optimize for physic calculation

[@dz-tech19]

for benchmarking just use some linux cpu load scripts . not sure if it will work properly but the last time i worked with a preconfigured one of those machines the scripts that uses multiprocessing orked fine on all cores of the cpus . it was 7 or 8 years ago . so i am not realy sure what was on those machines all i can remember is that they are using mpi to distrubut the load over the whole system and i tested a small script and it worked well

[@ayoubharry]

i'm watching the cat and i'm like did he overclock his cat lol

[@jlelelr]

can you run SuperPhi?

[@jesset2550]

I actually have a xeon phi it works on threadripper and ryzen

[@axe863]

How did you get it to work?

[@bachirkallas]

Where i can find xeon phi Windows 10 driver?Thanks

[@DiyintheGhetto]

Hello, Is it hard to setup one of these pcie cards?

[@petergriffin760]

do you think chained cooling is worth taking 600-800W power in single tube? no chance for separate paths for each coprocessor and CPU?

[@der8auer]

You can't overclock those CPUs. They won't care if the core is running at 75°C or 80°C

[@petergriffin760]

@@der8auer hey, i realy have no experience with such devices and massive cooling system, but could you post temperature readings of each 4 chips? I mean, under nominal load? temperature gradient over tube path might be so huge, starting with lower temp for first and going up through last

[@maintenancemaintenance7484]

where is your GN Tool kit?

[@kaiserped]

The notification bell was weird for me. I clicked on the English one and the video is in Deutsch. The German one is English video.

[@der8auer]

yea I mixed up the title

[@raphaelmphiri1891]

What's the name of the casing you used?

[@KapilDevMishra]

does the standard linux distro can utilize this?

[@charleshines6155]

The passive coolers on those are probably similar to that used on the Tesla cards. Tesla (not the car or famous inventor Nikola Tesla) are usually VERY expensive, even used ones can be far out of most people's price range. I am thinking that what ever I see now is probably doable on a Tesla card. Some Tesla cards do have fans but many of them are passive cooled.

[@Bojcha76]

Do we already know what is this pink thermal paste is?

[@Monasucks]

600W TDP alone from the Phi's plus maybe another 300+ by the CPUs and another 100 from the rest...

[@BinkyTheToaster]

At least your cat likes it, that's something.

[@kozad86]

I wish there was a way to use them for a VM.

[@hjq0023]

Can this co processor be installed in a gaming PC?

[@Rainbow__cookie]

i like that the 2 noctua cpu coolers look like a big ass cooler

[@phanb]

Hi @der8auer any updates on this project?

[@wino0000006]

Don't they keep that workstations in a separate rooms with air conditioning like it's done for servers?

[@dier2tte]

Pressed on english title ... german video. Pressed o german title ... finally english video 😄 Nice Roman 😄

[@wewillrockyou1986]

Cool stuff. Was just wondering, what tubing are you using for the water loop?

[@der8auer]

EK ZMT