A Mini Supercomputer Under 1500$

>32 cores, 64 threads
>64 GiB RAM
>1449.67$
1) Why aren't people doing this
2) What tasks can this actually do (compiling? video editing? mining monero? being a normal server?)

Attached: qwe.png (4312x1572, 2.12M)

Other urls found in this thread:

en.wikipedia.org/wiki/TOP500
en.wikipedia.org/wiki/Sunway_SW26010
en.wikipedia.org/wiki/Summit_(supercomputer)
twitter.com/SFWRedditImages

#2 answers #1. You can get a computer that is sufficient for most tasks for $1000 less. Why would people spend that much money on a computer for no reason?

Would rather have a NVidia Jetson Xavier desu

well, people buy threadrippers or even i9s because they apparently need those cores. a threadripper 1950x build will cost more than 1500$, yet will offer only half the cores.

>dual channel memory

Just get a full tower and put a crossfire of vega 64, 20 teraflops easy.
For comparison.

The fastest supercomputers based on AthlonMP:

Rutgers University, Department of Physics & Astronomy. Machine: NOW Cluster—AMD Athlon. CPU: 512 AthlonMP (1.65 GHz). Rmax: 794 GFLOPS.

Is this pic implying that 1 set of those parts is $360? new ryzen 7 alone is just a bit less than that

this is a build that focuses on cheap modern cores so I picked the 1700
make that 8 :^)
yeah but GPUs can't compile or function as servers yet.

yes this is called a beowulf cluster
it's only really useful if you have software that can use mpi or has their own distributed communication system like distcc. render farms, many niche HPC applications like NAMD, GROMACS, some FEM software, compilation would all work well, but without this or similar interfaces, most conventional software wouldn't be able to run in this scalable format regardless of how parallelized it is. it's also ridiculously costly to communicate from one node to another so the more embarrassingly parallel the workload is and stronger the network bandwidth is, the better. however, ryzen has neutered AVX so you'll lose performance on certain HPC applications.

however, if the code offers cuda/opencl codepaths (more relevant for cuda as amd opencl libraries as well as their compiler is broken, hence rocm) it is much more scalable with gpus - 8.5TFLOPS with a vega 56 if you could get it at ~450 MSRP where a 300USD 2700X won't be even close to that (i would wager 400GFLOPS if you could saturate its 6-wide design)

i would if i had the money
im waiting for 7nm for better thermals and clocks (have a 2700x until then)
ram prices are fucked, waiting for good coolers for TR4

>Why aren't people doing this
Nevermind me,
I'm just using my noisy, heat emitting, power hungry 1000watts machine just to surf the internets and watch movies.

Attached: 1505480354467.jpg (350x322, 13K)

>64 GiB RAM
>server
haha no

You forgot the cost of a switch between those 4, ssd/hdd x4 gfx x 1 at least.

I'm pretty sure you come out ahead with one system. cost of a switch, 3x hdd go away, as does the bottleneck. 4x 400W psu is gonna be more wasteful than 1x 1500W PSU as well, so electrical and space costs should be taken into account.

Wait let me get this straight, you want to daisy chain 4 shitty budget ryzens as a single pc and you didn't even buy a GPU or storage?
Even if we were talking about a render farm those nodes are stupid.

Clustering them is kind of slow over Ethernet, it's also much less convenient to manage a Beowulf cluster instead of a single workstation, and most workstation software won't run on a cluster
Also, a lot of the software that runs well on clusters will run better on, and if you will only run HPC stuff see The only use for a cluster like your idra for most Jow Forumsfags it's compiling with distcc, mining monero and ghetto video encoding, maybe BOINC and F@H but almost no user runs that
You don't need an HDD to run them, you can just boot them from PXE and run the OS from RAM
You do need a GPU with desktop Ryzen just to POST iirc

Can I get an actual parts list on this?

how do

Attached: 1510880855985.png (206x222, 63K)

thats what u get for buying a i9 i-i mean xeon

What ram is that?

>asrock

Attached: 1304376955947.png (600x450, 457K)

A supercomputer is a computer with a high level of performance compared to a general-purpose computer.

You'll never get to compute on a supercomputer, nor will you ever have anything that big/important to compute.

>What is Amazon AWS?

Until that cheap power supply blows up and takes the motherboard, CPU, and RAM with it.

>itx
>95w tdp
good luck

Attached: feelin hot hot hot.jpg (400x259, 42K)

Cute user, throw a SAN at it and you've got a nice little cluster. Could easily run vSphere or some other kind of converged software defined infrastructure hypervisor
Very easy to scale too Vs something like a threadripper build if all you want is stable compute.
Realistically only worthwhile for learning as that level of stability warrants renting data centre space if you'll just be sticking on a home internet connection.

Attached: FB_IMG_1525612305575.jpg (750x730, 38K)

>nevermind me
>one word
You want me to shove a Nirvana album up your ass?

we out here playin minesweeper and shit

Attached: minesweeper.png (1024x768, 120K)

I have this exact build actually. Use it for VR and as my work from home machine (software). -j16 compile times under clang 7 are about 3 times faster than -j4 on my MacBook pro. It runs any game I care to throw at it. Quiet and small.
i like it/10

touch for scale

Attached: IMG_20180318_165246.jpg (2688x1512, 1.75M)

the us military bought shitloads of ps3s

what case is that? it looks cool and has a lot of airflow

>-j16 compile times under clang 7 are about 3 times faster than -j4 on my MacBook pro.
This doesn't sound good.

yep. rendering uses those cores easily. the only time you wouldn't need those cores is if you had other shit to do while your shit renders.
I run a 1700X at 3.9ghz and it takes about 30 minutes to do a 2 pass h264 1080p at slower preset. which is the best quality without going into placebo. bitrate is irreverent since it doesn't really change encode times.
that being said.. if someone is doing a 4K render it probably takes them an hour on a 1700X or more. and if it cuts into their edit or work time they'd definitely want a better machine.

>Supercomputer
>Super
>AMD

en.wikipedia.org/wiki/TOP500

Titan uses opteron when they were bad, yet still is in the top.

en.wikipedia.org/wiki/Sunway_SW26010

INTEL & AMD ARE FINISHED & BANKRUPT

Not anymore sweetie
en.wikipedia.org/wiki/Summit_(supercomputer)

>GPUs can't compile
I wrote an SML compiler in OpenCL few years ago for fun. I'm pretty sure you could do the same for C or any other language.

If the Macbook Pro is able to turbo it's two cores to a similar speed as the R7 1700 (which is possible, since the 1700 all core boosts to 3.2GHz) this kind of scaling is reasonable. Bare in mind that some of the time running make will be spent linking rather than compiling, which won't see much of a speedup from throwing more cores at it. All things considered, 3x scaling is kind of what I'd expect.

What are these actual parts?

Where the fuck do you find memory that cheap?

ncase m1

actually lists the parts and prices or some sources faggot, not just a picture with products.

>amd
more like supercomPOOter rajeesh

That motherboard is complete and absolute garbage.
Avoid that one at all cost.
Hell, even the B350m it's shit, look it out.
Constant freezes and crashes, it kills SSDs and GPUs.

Attached: spesurprised.png (1280x720, 521K)

Or you could build one of those, buy a $800 1080ti, then actually have a usable computer.

How do you chain them to act as one computer?

Attached: 1524473090256.jpg (221x250, 6K)

special software like distcc that shares the load between them.

What about winblows? How do you do it?

i'd rather not share a load between them
>>/sqg/

>garbage
Expllain please i wanted A320M-hdv for my budget bulid with 2200G

but i thought Jow Forums loved assrock

Number 2. Rendering Farm.

>buy ES Xeon from Ebay
>buy motherboard and psu from case-builder with more money than sense
>buy gpu from kid who now plays minecraft on 2x1080ti
>buy ram from store at roughly the price of the rest of the build
cheapest damn computer I ever built

Attached: totally_legit.png (403x402, 24K)

>You do need a GPU with desktop Ryzen just to POST iirc

Any source on that? I got an athlon x4 845 headless setup as a router and it boots fine. Did they change that for the Ryzen line?

Any decent highly parallelized code runs on GPUs nowadays

The better option is a server board or bitcoin mining board and a fat stack of 1080's if you want to do simulations, etc.

mind = blown

you had bad experience with it? care to share?

Attached: 1527144567164.jpg (640x775, 106K)

>>/sqg/

What this mean?