Help on obtaining computational power

[PBH]

Hi guys, I'm sorry if this is not a relevant topic for this forum as I am new here. I'll explain my problem as follows.

Currently I'm working on my PhD right now which is partially funded through a government grant and am having difficulties buying computational hardware.

1. I am using classical molecular dynamics (mainly) to study one dimensional semiconductor nano-materials for thermoelectric applications. I am mostly looking into Si nano-wires and thermal conductivity suppression through diameter modulation. I'll provide details on exactly what I'm doing if required.

2. I am in Sri-Lanka and just started the PhD (I am the first student studying this field from the department although my supervisor followed the same field and is guiding me through the process).

3. Currently, I am using a workstation available in the department Intel Xeon-W2155 (32GB ram) with quadro P620 graphics card and also a PC which I built through my own expenses (i9 9900K without a GPU).

4. I am currently being paid a stipend sufficient to account for my living costs through a grant.

5. However, as I keep working on the simulations, I now understand that I am seriously being limited by the available computational power.

6. I am also saving money to buy a Quadro RTX 4000 (which is the maximum I can go for) but would probably have to wait for another 5-6 months until I can afford it.

7. Would anyone please consider if it would be possible to provide me with some sort of assistance in obtaining better computational capabilities (CPU based or GPU based both works). Either by providing remote access to perform the simulations or by providing me with some decent equipment which may be getting discarded ( I am willing to pay a reasonable amount along with shipping and wouldn't mind trying to repair a broken graphics card even if any are available to be spared).

8. I am open for suggestions as well and would be able to provide a written document explaining the situation through the head of the department if required.

9. I am also open to acknowledging any assistance in research articles which may be published.

 

I would very much appreciate it if someone considers assisting me on this to some extent. Also, I understand that excess computational power is not something that is freely available and also that what I am asking for may be way too much to be provided.

Thank you

[Grabhanem]

6 minutes ago, Buddima said:

Hi guys, I'm sorry if this is not a relevant topic for this forum as I am new here. I'll explain my problem as follows.

Currently I'm working on my PhD right now which is partially funded through a government grant and am having difficulties buying computational hardware.

1. I am using classical molecular dynamics (mainly) to study one dimensional semiconductor nano-materials for thermoelectric applications. I am mostly looking into Si nano-wires and thermal conductivity suppression through diameter modulation. I'll provide details on exactly what I'm doing if required.

2. I am in Sri-Lanka and just started the PhD (I am the first student studying this field from the department although my supervisor followed the same field and is guiding me through the process).

3. Currently, I am using a workstation available in the department Intel Xeon-W2155 (32GB ram) with quadro P620 graphics card and also a PC which I built through my own expenses (i9 9900K without a GPU).

4. I am currently being paid a stipend sufficient to account for my living costs through a grant.

5. However, as I keep working on the simulations, I now understand that I am seriously being limited by the available computational power.

6. I am also saving money to buy a Quadro RTX 4000 (which is the maximum I can go for) but would probably have to wait for another 5-6 months until I can afford it.

7. Would anyone please consider if it would be possible to provide me with some sort of assistance in obtaining better computational capabilities (CPU based or GPU based both works). Either by providing remote access to perform the simulations or by providing me with some decent equipment which may be getting discarded ( I am willing to pay a reasonable amount along with shipping and wouldn't mind trying to repair a broken graphics card even if any are available to be spared).

8. I am open for suggestions as well and would be able to provide a written document explaining the situation through the head of the department if required.

9. I am also open to acknowledging any assistance in research articles which may be published.

 

I would very much appreciate it if someone considers assisting me on this to some extent. Also, I understand that excess computational power is not something that is freely available and also that what I am asking for may be way too much to be provided.

Thank you

What type of compute do you need (FP32, FP16, int32, FP64)? Quadro RTX cards have good 16 and 32 bit efficiency but do pretty badly on FP64, which is what a lot of scientific workloads use.

If you're looking for a cheap-ish FP64 card, you can often find old Tesla compute cards on the used market for much less than an RTX 4000 that will absolutely flatten it in FP64. Just be aware that you may have to rig up a fan somehow as many of these server cards don't have active cooling.

 

I believe older AMD cards are also pretty decent in FP64 if you can find a good price on one used, but they don't touch a Tesla K40 or similar in that regard.

[PBH]

35 minutes ago, Grabhanem said:

What type of compute do you need (FP32, FP16, int32, FP64)? Quadro RTX cards have good 16 and 32 bit efficiency but do pretty badly on FP64, which is what a lot of scientific workloads use.

If you're looking for a cheap-ish FP64 card, you can often find old Tesla compute cards on the used market for much less than an RTX 4000 that will absolutely flatten it in FP64. Just be aware that you may have to rig up a fan somehow as many of these server cards don't have active cooling.

 

I believe older AMD cards are also pretty decent in FP64 if you can find a good price on one used, but they don't touch a Tesla K40 or similar in that regard.

Hi, thank you for the reply. Currently I am using mixed precision where both FP64 and FP32 are being used. Yes, it would be ideal to go for FP64, however with current resources I have been using the mixed option where the computations being done are reasonably accurate since my systems are homogeneous. Hence I am looking more towards FP32 performance.

[Letgomyleghoe]

< content removed >

 

i don’t know if this is out of budget, but is there any super computer time slots you could rent out?

Edited March 29, 2020 by LogicalDrm

[minibois]

You could maybe look into making a BOINC project, so people could make use of their own hardware to contribute to your cause.

I would be weary of someone who I do not know accessing my computer from a distance.

[straight_stewie]

2 hours ago, Buddima said:

1. I am using classical molecular dynamics (mainly) to study one dimensional semiconductor nano-materials for thermoelectric applications. I am mostly looking into Si nano-wires and thermal conductivity suppression through diameter modulation. I'll provide details on exactly what I'm doing if required.

2. I am in Sri-Lanka and just started the PhD (I am the first student studying this field from the department although my supervisor followed the same field and is guiding me through the process).

Have you considered a service like Amazon AWS EC2?

Remote compute power is fairly cheap. A machine with 16 K80 GPUs and 64 Broadwell Xeon Cores (p2.16XLarge) has a base cost of $14.00USD per hour, however this price can fluctuate significantly based on demand and fluctuation (usually the price goes down). This configuration should give roughly 70 TFLOPS single precision and 23 TFLOPS double precision performance.

AWS has many, many different machine configurations optimized for compute, either GPU accelerated or not, that come in at different price/performance ratios. Some are overwhelmingly cheap, coming in at a few cents per hour, although in terms of compute performance you get what you pay for, so you need to do an analysis to figure out which will be the cheapest overall.

[PBH]

15 minutes ago, minibois said:

You could maybe look into making a BOINC project, so people could make use of their own hardware to contribute to your cause.

I would be weary of someone who I do not know accessing my computer from a distance.

Thank you. I shall look into that. Really appreciate it guys

 

7 minutes ago, straight_stewie said:

Have you considered a service like Amazon AWS EC2?

Remote compute power is fairly cheap. A machine with 16 K80 GPUs and 64 Broadwell Xeon Cores (p2.16XLarge) has a base cost of $14.00USD per hour, however this price can fluctuate significantly based on demand and fluctuation (usually the price goes down). This configuration should give roughly 70 TFLOPS single precision and 23 TFLOPS double precision performance.

AWS has many, many different machine configurations optimized for compute, either GPU accelerated or not, that come in at different price/performance ratios.

I had not considered that. Thanks

 

16 minutes ago, scuff gang said:

I don’t think they were begging, they were asking if someone would be able to allow remote access to perform simulations for there phd, or if they had used hardware he could possibly buy.

 

 

i don’t know if this is out of budget, but is there any super computer time slots you could rent out?

There is no one that I know of as of yet. Except for AWS as mentioned by Stewie in the previous reply.

[LogicalDrm]

Moderators note:

This seems to be good to leave open for realistic solutions. Asking to access others computers is not one of them. Anonymous distributed computing (what I think they mean by BOINC) is more realistic, so is buying time from some supercomputer (your instructor should know which universities or research companies are good to contact).

[PBH]

Thanks for all the helpful tips guys. I found a used Tesla K40 on amazon for 1/4th the price of an RTX 4000. I'm also looking into the services being offered by AWS and BOINC. Just a few small concerns though. In any case if I try to buy and install a K40 to my system, would the mounting be any different from that of a regular desktop graphics card (say a quadro P620 or RTX 2080)? Also what would be considered as safe temperatures for a K40 to run under. During the night time and weekends, all air conditioning in the university will be shut down and the ambient temperature would be close to 33-35 Celsius usually, while I plan to use the system 24x7. I am hoping to attach an exhaust fan from the rear end of the card to suck air through the heat sink (but I have no idea regarding the required flow rate) or to remove the plastic housing and attach two small fans to blow directly onto the heat sink as done in gaming GPUs. Which should be the more reliable and recommendable option. I have never physically seen a server grade system so am not familiar with the mounting mechanisms or thermal solutions except for the knowledge that they are typically passively cooled.

 

Thank you

[Pickles von Brine]

On 3/30/2020 at 2:10 AM, Buddima said:

Thanks for all the helpful tips guys. I found a used Tesla K40 on amazon for 1/4th the price of an RTX 4000. I'm also looking into the services being offered by AWS and BOINC. Just a few small concerns though. In any case if I try to buy and install a K40 to my system, would the mounting be any different from that of a regular desktop graphics card (say a quadro P620 or RTX 2080)? Also what would be considered as safe temperatures for a K40 to run under. During the night time and weekends, all air conditioning in the university will be shut down and the ambient temperature would be close to 33-35 Celsius usually, while I plan to use the system 24x7. I am hoping to attach an exhaust fan from the rear end of the card to suck air through the heat sink (but I have no idea regarding the required flow rate) or to remove the plastic housing and attach two small fans to blow directly onto the heat sink as done in gaming GPUs. Which should be the more reliable and recommendable option. I have never physically seen a server grade system so am not familiar with the mounting mechanisms or thermal solutions except for the knowledge that they are typically passively cooled.

 

Thank you

The cards use chassis fans to push air through them at VERY high rates. We are talking 200-400+ CFM at full load. So, a normal fan will not keep it cool. Plus, you will need a second video card to use as display out. As far as everything else goes it is just a standard card. Still uses the ATX mounting system. My worry is just keeping it cool. Honestly, to me, it is not worth the hassle. You could probably rent time from AWS for what you need. 

https://aws.amazon.com/emr/pricing/

[PBH]

2 hours ago, Pickles - Lord of the Jar said:

The cards use chassis fans to push air through them at VERY high rates. We are talking 200-400+ CFM at full load. So, a normal fan will not keep it cool. Plus, you will need a second video card to use as display out. As far as everything else goes it is just a standard card. Still uses the ATX mounting system. My worry is just keeping it cool. Honestly, to me, it is not worth the hassle. You could probably rent time from AWS for what you need. 

https://aws.amazon.com/emr/pricing/

Thank you. I think I now have all the information I needed. Very much appreciate the helpful comments.

[Pickles von Brine]

Just now, Buddima said:

Thank you. I think I now have all the information I needed. Very much appreciate the helpful comments.

Anytime!