PrimeGrid
Please visit donation page to help the project cover running costs for this month

Toggle Menu

Join PrimeGrid

Returning Participants

Community

Leader Boards

Results

Other

drummers-lowrise

Advanced search

Message boards : Number crunching : Bang for the buck?

Author Message
Profile composite
Volunteer tester
Send message
Joined: 16 Feb 10
Posts: 841
ID: 55391
Credit: 782,673,526
RAC: 442,633
Discovered 2 mega primesFound 1 prime in the 2018 Tour de Primes321 LLR Turquoise: Earned 5,000,000 credits (5,477,467)Cullen LLR Gold: Earned 500,000 credits (776,297)ESP LLR Ruby: Earned 2,000,000 credits (3,433,680)Generalized Cullen/Woodall LLR Ruby: Earned 2,000,000 credits (2,093,491)PPS LLR Sapphire: Earned 20,000,000 credits (27,299,451)PSP LLR Turquoise: Earned 5,000,000 credits (6,587,988)SoB LLR Sapphire: Earned 20,000,000 credits (33,929,863)SR5 LLR Turquoise: Earned 5,000,000 credits (6,110,877)SGS LLR Ruby: Earned 2,000,000 credits (3,486,285)TRP LLR Turquoise: Earned 5,000,000 credits (7,070,795)Woodall LLR Amethyst: Earned 1,000,000 credits (1,693,614)321 Sieve (suspended) Emerald: Earned 50,000,000 credits (50,256,050)Cullen/Woodall Sieve (suspended) Turquoise: Earned 5,000,000 credits (5,571,178)Generalized Cullen/Woodall Sieve (suspended) Emerald: Earned 50,000,000 credits (50,009,610)PPS Sieve Double Silver: Earned 200,000,000 credits (384,567,672)Sierpinski (ESP/PSP/SoB) Sieve (suspended) Jade: Earned 10,000,000 credits (10,165,888)TRP Sieve (suspended) Sapphire: Earned 20,000,000 credits (20,071,454)AP 26/27 Turquoise: Earned 5,000,000 credits (6,616,128)GFN Emerald: Earned 50,000,000 credits (53,771,465)WW Gold: Earned 500,000 credits (932,000)PSA Double Bronze: Earned 100,000,000 credits (102,762,384)
Message 150117 - Posted: 17 Apr 2021 | 5:11:01 UTC

Would it make financial sense to crowdfund the purchase and operation of a datacenter-scale GPU dedicated to PrimeGrid? Not especially advocating this particular brand which has 4 NVidia A100 GPUs for US$65,000 (there are other brands), compared to buying and running a personal high end NVidia GPU and computer to feed it.

Would it be feasible to divide up the subproject credit (and prime reporting) to the funders as single computing resource, or use it as a virtual GPU for each of the funders?

Would there be resale value in the contributions, i.e. a secondary market for the machine fraction?

Profile composite
Volunteer tester
Send message
Joined: 16 Feb 10
Posts: 841
ID: 55391
Credit: 782,673,526
RAC: 442,633
Discovered 2 mega primesFound 1 prime in the 2018 Tour de Primes321 LLR Turquoise: Earned 5,000,000 credits (5,477,467)Cullen LLR Gold: Earned 500,000 credits (776,297)ESP LLR Ruby: Earned 2,000,000 credits (3,433,680)Generalized Cullen/Woodall LLR Ruby: Earned 2,000,000 credits (2,093,491)PPS LLR Sapphire: Earned 20,000,000 credits (27,299,451)PSP LLR Turquoise: Earned 5,000,000 credits (6,587,988)SoB LLR Sapphire: Earned 20,000,000 credits (33,929,863)SR5 LLR Turquoise: Earned 5,000,000 credits (6,110,877)SGS LLR Ruby: Earned 2,000,000 credits (3,486,285)TRP LLR Turquoise: Earned 5,000,000 credits (7,070,795)Woodall LLR Amethyst: Earned 1,000,000 credits (1,693,614)321 Sieve (suspended) Emerald: Earned 50,000,000 credits (50,256,050)Cullen/Woodall Sieve (suspended) Turquoise: Earned 5,000,000 credits (5,571,178)Generalized Cullen/Woodall Sieve (suspended) Emerald: Earned 50,000,000 credits (50,009,610)PPS Sieve Double Silver: Earned 200,000,000 credits (384,567,672)Sierpinski (ESP/PSP/SoB) Sieve (suspended) Jade: Earned 10,000,000 credits (10,165,888)TRP Sieve (suspended) Sapphire: Earned 20,000,000 credits (20,071,454)AP 26/27 Turquoise: Earned 5,000,000 credits (6,616,128)GFN Emerald: Earned 50,000,000 credits (53,771,465)WW Gold: Earned 500,000 credits (932,000)PSA Double Bronze: Earned 100,000,000 credits (102,762,384)
Message 150120 - Posted: 17 Apr 2021 | 16:34:15 UTC
Last modified: 17 Apr 2021 | 16:40:50 UTC

Benchmarks are required to make an assessment.

A100 performs similar to RTX3090 when it's processing bandwidth-limited computations like FFT.

For example, this series of benchmarks doesn't show huge throughput improvement over RTX3090 when the A100 is sitting on a PCIe bus for some workloads.

Profile composite
Volunteer tester
Send message
Joined: 16 Feb 10
Posts: 841
ID: 55391
Credit: 782,673,526
RAC: 442,633
Discovered 2 mega primesFound 1 prime in the 2018 Tour de Primes321 LLR Turquoise: Earned 5,000,000 credits (5,477,467)Cullen LLR Gold: Earned 500,000 credits (776,297)ESP LLR Ruby: Earned 2,000,000 credits (3,433,680)Generalized Cullen/Woodall LLR Ruby: Earned 2,000,000 credits (2,093,491)PPS LLR Sapphire: Earned 20,000,000 credits (27,299,451)PSP LLR Turquoise: Earned 5,000,000 credits (6,587,988)SoB LLR Sapphire: Earned 20,000,000 credits (33,929,863)SR5 LLR Turquoise: Earned 5,000,000 credits (6,110,877)SGS LLR Ruby: Earned 2,000,000 credits (3,486,285)TRP LLR Turquoise: Earned 5,000,000 credits (7,070,795)Woodall LLR Amethyst: Earned 1,000,000 credits (1,693,614)321 Sieve (suspended) Emerald: Earned 50,000,000 credits (50,256,050)Cullen/Woodall Sieve (suspended) Turquoise: Earned 5,000,000 credits (5,571,178)Generalized Cullen/Woodall Sieve (suspended) Emerald: Earned 50,000,000 credits (50,009,610)PPS Sieve Double Silver: Earned 200,000,000 credits (384,567,672)Sierpinski (ESP/PSP/SoB) Sieve (suspended) Jade: Earned 10,000,000 credits (10,165,888)TRP Sieve (suspended) Sapphire: Earned 20,000,000 credits (20,071,454)AP 26/27 Turquoise: Earned 5,000,000 credits (6,616,128)GFN Emerald: Earned 50,000,000 credits (53,771,465)WW Gold: Earned 500,000 credits (932,000)PSA Double Bronze: Earned 100,000,000 credits (102,762,384)
Message 150133 - Posted: 18 Apr 2021 | 14:13:04 UTC

Let's be more clear about this.

The question is always rent or buy if you're not using the surplus capacity of some other owner's GPU farm. Six years down the road your RTX 3090 is going to look as dated as the GTX 760 I'm currently running, and you'll have spent an average of $300 per year on the acquisition cost alone.

NVidia's newest CPU+GPU line "Grace" is data-center only. It won't be the end-of-the-road for consumer-grade GPUs for a while, just as chip makers still make enthusiast-level CPUs. Their price however is going through the roof as there will be stiff competition for dwindling supply. GPU volume is shifting to mobile and cloud as AI is ramping up to become the dominant application for GPUs. Sooner or later the rent-or-buy question will be about datacenter-based GPU, whether it's in the public cloud or some organization's compute resource.

The largest contributors to PG on GPU subprojects are either already renting cloud-based GPUs or they are relying on fleets of PCs set up for business or education. The fleets will evaporate in the next round of hardware refresh as computing moves to the cloud at scale simply because of datacenter economics.

So at some point PrimeGrid members who want to own GPU capacity will have to decide whether to spend ever more money on a top-of-the-line consumer-grade GPU, or a buy a share of a datacenter-based resource. The question hasn't been broached before, but this is certain to happen within the expected lifetime of many GPU subprojects.

Yves Gallot
Volunteer developer
Project scientist
Send message
Joined: 19 Aug 12
Posts: 672
ID: 164101
Credit: 305,042,960
RAC: 0
GFN Double Silver: Earned 200,000,000 credits (305,042,960)
Message 150134 - Posted: 18 Apr 2021 | 15:24:09 UTC

Cloud-based computations is the future, probably but on what sort of hardware?

AI (I prefer to say Deep Learning because it is not logic and then not intelligent) "GPUs" operate on 16-bit floats or less. Today no fast algorithm exists for multiplying large numbers with this type of data.

GPUs for high-performance computing (NVidia A100 and 64-bit floats) are fast for specific algorithms that can't be distributed efficiently on many CPUs. But PrimeGrid computation is a large number of independent tests then few dozen Intel CPUs is certainly more efficient.

The trick with GeForce series is to use shader cores for the calculations. The types are 32-bit integers or floats but because of the large number of cores, GPUs outperform CPUs especially for parallel tasks like sieve or the WW project. Graphic GPUs can also be used to check primality of large numbers but then a Number Theoretic Transform (based on Int32) must be used rather than a Fast Fourier Transform (based on Float64).

Then today a fast computer for PrimeGrid is a gaming desktop rather than an expensive computer dedicated to high-performance computing. Then cloud gaming platforms should be a cheapest choice. The problem is to hide the computation in a game, but miners have probably solved it.

Post to thread

Message boards : Number crunching : Bang for the buck?

[Return to PrimeGrid main page]
DNS Powered by DNSEXIT.COM
Copyright © 2005 - 2021 Rytis Slatkevičius (contact) and PrimeGrid community. Server load 0.90, 1.52, 1.69
Generated 18 Sep 2021 | 8:44:53 UTC