External GPU's

[Brucifer]

Has anyone here looked into using external gpu's for boinc crunching ?????

[Dirk Broer]

Not yet, I'm more into internal GPUs. Love to see my ARM cores using their GPU too...

[KarmannGaz]

What a great idea, I didn't know there was such a thing, but have just been reading this about them - https://arstechnica.com/gaming/2016/03/amd-wants-to-standardise-the-external-gpu/
(https://arstechnica.com/gaming/2016/03/amd-wants-to-standardise-the-external-gpu/)
Will have to do a bit more research, this might be a way to give my older machines a bit more life perhaps.

[NeoGen]

From the bits and pieces I read about it, it seems this latest generation are mostly Thunderbolt 3 connected enclosures, but the problem is that they are pretty much vendor locked-in. There are a handful of manufacturers and each only guarantees compatibility with their own top-of-the-line xyz laptop. I have not heard about an open solution that could be used with any Thunderbolt 3 desktop/laptop. And they're pretty expensive for just being a custom sized case with a Thunderbolt 3 to PCIe converter, the manufacturers are taking a ton of profit for that niche gamers market.

Bitcoin/Altcoin GPU miners have a much cheaper and effective solution like this:

http://blog.whitesites.com/blogs/large/Bitcoin-Mining-Rig-with-GPUs_635096703769853515.jpg

This one is an open case but I'm sure there's variations of this with closed case, I think I've seen plexiglass ones. And unlike the GPU enclosures that need a separate PSU for each enclosure, this will take all those GPUs with the same one PSU.

[NeoGen]

With a little more digging around on the subject I found that several of the custom cases out there that hold 6 or even 8 GPUs in a standard 4U server rack mountable case connect their GPUs using these adapters:

https://i.stack.imgur.com/Pg5L0.jpg

It seems they're USB to PCIe 3.0 powered risers / adapters. This I believe will severely cut the bandwidth of the GPUs full x16 (128Gbps) to x1 (8Gbps) only, and on top of that it even takes another cut down to 5Gbps because that's the limit of USB 3.0. But I guess for GPU mining the need for data transfer back and forth is minimal compared to the need for computational capacity so the bandwidth is really not that necessary. In terms of Distributed Computing it might work too, but I think some projects might get a performance penalty due to the higher latency and lower bandwidth of USB 3.0.

Who's got $500 to splurge (http://www.miningrigs.net/?product=gray-matter-gpu-server-case-v3-0-2) (plus cost of GPUs, PSU, motherboard, RAM, CPU, etc) to test it out? :icon_mrgreen: :icon_lol:

[Brucifer]

Well I was looking at it from the aspect of crunching large long term work units for primegrid. and I wouldn't think that the band width would be an issue in something like that as compared to the gamers that are driving tons of video through the system..... thoughts?? And also to help spread stuff out to help in the cooling issue too.

[NeoGen]

Well I was looking at it from the aspect of crunching large long term work units for primegrid. and I wouldn't think that the band width would be an issue in something like that as compared to the gamers that are driving tons of video through the system..... thoughts?? And also to help spread stuff out to help in the cooling issue too.
I've thought about one of those GPU mining solutions but for crunching, but I don't have the funds for it yet. :icon_rolleyes:
I do think for long workunits it would do perfectly, and most projects have medium to long workunits that don't require CPU feeding so it would be awesome. But if you get a big rig like the ones above remember it will produce a massive amount of heat. I have a small "home office" (man cave :icon_wink: ) and just running one good GPU and a couple of desktops raises the temperature of this room considerably compared to the rest of the house, I can't imagine how I would run a monster 6-GPU cruncher.

[Dirk Broer]

It might both be cheaper and cooler to hire GPUs on the Amazon Cloud...

[NeoGen]

Their prices are confusing being broken down by regions, and then depending on the instance type and etc, but from a quick look up they still seem way too high... the lowest I could tell for a GPU powered unit was $0.9 per hour, which is crazy high since we want to run it 24/7.

Now that I think of it... Bitcoin and similars have "cloud mining" which is you rent a big server to mine for you and (if the price is good) it will turn you a small profit. In the case of Distributed Computing the reward is in BOINC cobblestones and/or other kinds of points that have no monetary value, so I'm looking at $0.9/hour as crazy because I'm not getting that money back (and let's be honest... it's really way more than if I just buy my own GPU and pay for the electricity).
Now if there was a similar cloud service that would cost about... $40/month for one good GPU instance for BOINC crunching? I would totally jump on it! In my view $480/year is not too far from the cost of the GPU (well, some are way more than that) plus electricity to run it. :)

EDIT: You can tell by the numbers I put up that I don't buy the biggest baddest GPUs, right? :icon_lol:

[Nflight]

We have such a fascinating hobby... :blob3:

[Brucifer]

I'm not really into buying into the cloud stuff. If I'm going to be spending money I'd rather add it to my farm as at least the hardware is mine to use as long as I want. Maybe I'm missing something but thats just the way I feel.

[NeoGen]

I'm not really into buying into the cloud stuff. If I'm going to be spending money I'd rather add it to my farm as at least the hardware is mine to use as long as I want. Maybe I'm missing something but thats just the way I feel.
I feel you. I'm not biting into the cloud stuff either at current prices, but I'll be honest, if it was a reasonable price and it took away all the time I spend setting up, testing, monitoring, fixing, etc... I'd be on it. I'd still keep some physical one for myself to play with of course, but my majority would be on the cloud. But not at current prices.

[Dirk Broer]

I feel you. I'm not biting into the cloud stuff either at current prices, but I'll be honest, if it was a reasonable price and it took away all the time I spend setting up, testing, monitoring, fixing, etc... I'd be on it. I'd still keep some physical one for myself to play with of course, but my majority would be on the cloud. But not at current prices.

Isn't Amazon offering a so-called free service (https://aws.amazon.com/free/terms/?ft=nt) (for 12 months) too?

[NeoGen]

From what I read it's only for micro instances (1 slow CPU core?), I didn't see GPU instances for the free tier. I think vaughan has one of these, I can't find the thread now but I remember he gave us a link to one of his computers in a project that's an amazon instance. It was an interesting little instance but from the details it did not seem powerful at all. :(

[vaughan]

I have forgotten what processor I had for the free AWS instance. Since then I have paid for (too) many cores. Last month the invoice was pretty high so I've stopped using AWS temporarily. The instance I was using was a Xeon E5-2666 v3 @ 2.9GHz with 36 cores available to me.

[Nflight]

I found this in my hunt for the next big purchase, External Enclosures (https://gpunerd.com/external-gpu-buyers-guide-egpu). :blob3:

[Brucifer]

Thanks for thar link Nflight! :-)

[vaughan]

You could get one of these for just under AUD1k

https://www.pccasegear.com/products/39669/gigabyte-aorus-1070-gaming-box

[Nflight]

The External graphics Unit is not really what we are all looking for, is it? We are instead looking for a GPU Server Farm and just how many GPU's we can muster into a confined space for the least amount of Cash$$$.

First Link is: Images for GPU Server Farm 2017 (https://www.google.com/search?q=gpu+server+farm+2017+boinc&rlz=1C1CHZL_enUS750US750&tbm=isch&tbo=u&source=univ&sa=X&ved=0ahUKEwj1leK_mZ7YAhWOS98KHXa8A1IQsAQISQ&biw=1920&bih=949)

Second Link is: GPU Server 2017 BOINC (https://www.google.com/search?q=gpu+server++2017+boinc&rlz=1C1CHZL_enUS750US750&source=lnms&tbm=shop&sa=X&ved=0ahUKEwjksOz_m57YAhWkhOAKHVmfALcQ_AUIDSgE&biw=1920&bih=949)

[Dirk Broer]

The External graphics Unit is not really what we are all looking for, is it? We are instead looking for a GPU Server Farm and just how many GPU's we can muster into a confined space for the least amount of Cash$$$.

First Link is: Images for GPU Server Farm 2017 (https://www.google.com/search?q=gpu+server+farm+2017+boinc&rlz=1C1CHZL_enUS750US750&tbm=isch&tbo=u&source=univ&sa=X&ved=0ahUKEwj1leK_mZ7YAhWOS98KHXa8A1IQsAQISQ&biw=1920&bih=949)

Second Link is: GPU Server 2017 BOINC (https://www.google.com/search?q=gpu+server++2017+boinc&rlz=1C1CHZL_enUS750US750&source=lnms&tbm=shop&sa=X&ved=0ahUKEwjksOz_m57YAhWkhOAKHVmfALcQ_AUIDSgE&biw=1920&bih=949)

We should befriend Bitcoin miners and other crypto currency cashers. Once their toys -a nine GPu miner here- (https://gigazine.net/gsc_news/en/20171219-inno3d-p104-100-mining/) are no longer economical, they dump them...Still there's the power bill to reckon with

[Tweeder]

I had stumbled upon a site that was selling a 4U Server chassis with 8 GPUs crammed into it, but cannot find it anymore. Price was about US 3k for a completed box including the Rx 470s they were putting in it. I figured it could be built for cheaper. Especially after cryptos are done. Full disclosure I have been mining a little bit of Ether to help fund my next build.

I think an 8 GPU unit will definitely show upon the electrical bill, but on the plus size it will be a nice sound machine and space heater.

[Dirk Broer]

Pure economically seen GPUs are a disaster. I might want to try a Bristol Ridge 35 Watt TDP APU though, until Raven Ridge comes along.
If it -35 Watt Bristol Ridge- works well, I might buy a bunch of 320/350 boards to replace my FM1 and FM2 systems.

[NeoGen]

The evolution of the coin miners went more or less from CPU -> GPU -> FPGA -> ASIC.

In Distributed Computing we have already CPUs and GPUs, not sure about FPGAs though, although the blue team is launching some chips with FPGAs built-in lately.
Now ASICs are very specific processors for single task only (and do it incredibly well and efficiently!) Since Distributed Computing has such a wide array of different subjects and computations its impossible to build 1 ASIC chip that could accelerate all projects, but I could very well see an ASIC chip dedicated to prime number finding and testing.
Can you imagine how much PrimeGrid would skyrocket if someone developed an ASIC chip dedicated to prime number computation only? :icon_mrgreen:

[Dirk Broer]

Can you imagine how much PrimeGrid would skyrocket if someone developed an ASIC chip dedicated to prime number computation only? :icon_mrgreen:

You'll need an ASIC per application of course, but there are multi-USB hubs that bitcoin miners previously used...

[Nflight]

Building a GPU Server (http://www.solverworld.com/building-a-multiple-gpu-computer-for-grid-computing/) & Installing BOINC on a multi-GPu Server (http://www.solverworld.com/install-boinc-gpu/)

[Nflight]

Lowest Price GPU Server (https://www.avadirect.com/6GPU-Tower-Mining-Server/Configure/11400706?sgd=333d333d314d316d315d316&gclid=EAIaIQobChMIiN2k7Kew2AIVKLizCh0DrAgQEAEYASAB EgJoqvD_BwE) comes with options too:

EPYC GPU Server (https://www.avadirect.com/Supermicro-A-Server-4023S-TRT-Dual-AMD-EPYC-7000-series-SAS-SATA-4U-Rackmount-Tower-Server-Computer/Configure/11535865)with options: