[QUOTE=stars10250;392285]It was running CUDALucas at the time of the screenshot and that makes extensive use of DP. I get the same clock numbers in MSI Afterburner overclock software, ASUS GPU tweak software, and in CPU-Z. I've seen the clock jump down when I unload it by turning off the CUDALucas calculation, but as soon as I launch the program it jumps back up to full speed.[/QUOTE]

I am pretty sure LaurV's comment applies to Titan, but not the newer cards with the lower DP ratio. Instead of slowing the card down to run DP calcs, DP code runs at 1/32 of SP code vs an older ratio of 1/24. So, you get 1/32 of a faster clock!

I know little about GPU architecture but I grabbed this image from a document on the GK110 that I have in my GTX780Ti. The double-precision units are colored yellow and presumably CUDALucas makes use of these. Does CL make use of all the single-precision cores? It's hard to believe they are sitting idle with the card throwing off 250W.

Does anyone use Amazon Elastic Compute Cloud for their GIMPS LL work? I'm only interested in LL, and it seems like it could be done on an Amazon CPU or GPU. Toward the end of each month I pay a penalty for power use (California) and it would probably be cheaper to do my computing elsewhere. I'm curious about the cost to do a LL test of a decent size exponent (60M+) on Amazon hardware. I will probably try it but I'd be interested in hearing if anyone has done this.

[QUOTE=stars10250;395711]Does anyone use Amazon Elastic Compute Cloud for their GIMPS LL work?[/QUOTE]

Chris H. has. He was describing how to do it here: [url]http://mersenneforum.org/showthread.php?p=388009[/url]
If I understood what he said, I might have fired up an instance or 2 to do work in the 100M digit range.
It would be cheaper than buying and running a machine locally, especially with the 'all green power' premium I pay.

[QUOTE=stars10250;395711]Does anyone use Amazon Elastic Compute Cloud for their GIMPS LL work? I'm only interested in LL, and it seems like it could be done on an Amazon CPU or GPU. Toward the end of each month I pay a penalty for power use (California) and it would probably be cheaper to do my computing elsewhere. I'm curious about the cost to do a LL test of a decent size exponent (60M+) on Amazon hardware. I will probably try it but I'd be interested in hearing if anyone has done this.[/QUOTE]

I wouldn't recommend it, unless you use an EBS-backed spot instance with volume deletion on termination disabled (in case you get outbid, you can restore the mprime files). EC2 on-demand tends to be quite expensive.

C4 gives you about 0.2 GHz more base clock CPU, and the C4's turbo mode is probably useless with mprime keeping the processor hot. C3 is probably a better bet for throughput per dollar.

The old cg1.4xlarge instances available in us-east-1 are by far the best deal for trial factoring. One zone has spot instances for around \$0.15/hr. They offer far more TF throughput/\$ than g2.2xlarge. I don't know which does better at CUDALucas. You could try both.

You might be better served by renting \$5 virtual machines at Digital Ocean. I did that as experiment for Seventeen or Bust work, which also uses mprime. I found I would get 40 to 80% of a CPU core. It's the cheapest CPU power I've found.

[QUOTE=Uncwilly;395716]If I understood what he said, I might have fired up an instance or 2 to do work in the 100M digit range. It would be cheaper than buying and running a machine locally, especially with the 'all green power' premium I pay.[/QUOTE]

I wish you had given it a try -- it really is quite easy once you "make friends" with it. At the very least I recommend people who are interested sign up for AWS's "Free Tier" promo -- one year of a t1.micro instance for free (well, technically, two instances; one Linux and one Winblows although the storage limits tend to prevent you from running both at the same time). It's a great way to get familar with the environment with no risk (but you won't get much LL'ing out of it).

To put on the table though, I will always defer to Mark's counsel on AWS / EC2. He's got WAY more experience with it than I.

Actually, Mark... This brings up a thought: you had talked about possibly publishing your AMI for TF'ing. Perhaps now would be a good time...?

And, seperately, with regards to TF'ing, with the recently higher spot prices for cg1.4xlarge even in us-east-1d (sometimes over 15 cents an hour (gasp! :wink:)) the value proposition of g2.2xlarge may soon need to be re-evaluated.

[QUOTE=chalsall;395744]I wish you had given it a try -- it really is quite easy once you "make friends" with it. At the very least I recommend people who are interested sign up for AWS's "Free Tier" promo -- one year of a t1.micro instance for free (well, technically, two instances; one Linux and one Winblows although the storage limits tend to prevent you from running both at the same time). It's a great way to get familar with the environment with no risk (but you won't get much LL'ing out of it).[/quote]

I would recommend starting with the t2.micro. That gives you a little more CPU to play with and is also included in the free tier. Instead of being throttled after 2 seconds, you get a bucket of CPU minutes that is replenished 6 minutes per hour (kept for 24 hours). Much of our development and some of our production environment runs on t2's.

[quote]To put on the table though, I will always defer to Mark's counsel on AWS / EC2. He's got WAY more experience with it than I.[/quote]

There are a bunch of services I haven't used, but we do spend well into the five figures every month on AWS.

[quote]Actually, Mark... This brings up a thought: you had talked about possibly publishing your AMI for TF'ing. Perhaps now would be a good time...?[/quote]

I can probably make time this week or on the weekend. Amazon recently added a new feature to where can poll to see if your spot instance is [url=https://aws.amazon.com/blogs/aws/new-ec2-spot-instance-termination-notices/]about to be terminated[/url]. I want to implement support for that to unreserve unfinished assignments in mfloop.py. I may also build a script to unreserve mprime work, but I need to test that behaviour with GPU72 proxying.

Chris, how does one go about unreserving a GPU72 proxied request? Unreserve at GPU72 or PrimeNet? I usually use my CPUs for SoB, so I haven't played with that feature much.

[quote]
And, seperately, with regards to TF'ing, with the recently higher spot prices for cg1.4xlarge even in us-east-1d (sometimes over 15 cents an hour (gasp! :wink:)) the value proposition of g2.2xlarge may soon need to be re-evaluated.[/QUOTE]

I'll investigate that while I test.

I think I can make a user agree to the mprime license agreement in the user data to enable using mprime. It can then be downloaded and started.

[QUOTE=Mark Rose;395761]Chris, how does one go about unreserving a GPU72 proxied request? Unreserve at GPU72 or PrimeNet? I usually use my CPUs for SoB, so I haven't played with that feature much.[/QUOTE]

Is this a TF, P-1 or LL or DC assignment?

PM me an example assignment and I'll drill down on how best to handle this.

I'm currently running two LL tests on two 8-core Xeons on Microsoft's Azure platform. Setup was easy with a virtual machine. I picked basic machines because I don't need a lot of ram or SSD support. They tend to be slower but have a cheaper hourly cost. I get $200 of free money for my 30 day trial. So far it looks prohibitively expensive. I should use up my money in about 2 weeks and will report back then.

My computing trial is done. I burned through $200 of free Microsoft Azure money doing LL testing. I ran on two 8-core machines. One was a 2.2 GHz E5-2660 Xeon and the other was an AMD Opteron at 2.1 GHz. These were on the low-end of capabilities, but the upper end machines had features I didn't need for LL testing such as solid state hard drives, fast internet connection, etc. and they were very expensive hourly (over $4/hr verses $.61/hr, they have a funny formula for cpu compute time). So the final analysis works out that the two machines burned through $200 in a week and completed approximately 95% of 1 LL test (combined) with exponents of size 72M. For comparison, my desktop PC does this calculation, assuming hardware amortization over 2 years, for about $9 including the cost of electricity.

Setting up Azure and running prime95 was straightforward. You give them your info, select the hardware, and it builds a virtual machine that operates just like a PC running Windows Server 2012. I went to the internet, downloaded prime95, and was up and running like it was just another build. You can monitor your daily costs (compute, bandwidth, storage). I seemed to get about 92% of each 8 core machine. At the end I offloaded the save-files to be completed by my machine.

I can see MS making a profit but this seems a bit harsh at 22 times more expensive than a desktop pc. I know it's more reliable, and has all sorts of features, but for what we do it doesn't make sense. I thought cloud computing was the next big thing but it's doomed if this is what they charge. It makes me consider going into the hpc business and charging something reasonable.

I'm running on Amazon's elastic computing cloud EC2 now. They gave me 1 CPU for a year free. It's an intel xeon 2.6 ghz with 1 GB memory. I get 30 GB storage. It's similar to microsoft's setup with a virtual machine, windows server 2012. There was more to do with setup of an RSA encryption key though. On a 72M exponent it is giving 30 ms iteration times.