Farm news
I have just 1 or 2 of the Intel GPU machines crunching overnight at the moment. The Parallella's and Raspberry Pi's are still running 24/7 though.
I have ordered a couple of new 6 core/12 thread machines to replace my oldest i7's which are also 6 core/12 thread. As the new machines will be running GTX750Ti's I have taken them out of the dual GPU cruncher and given them to the shop for the first build and ordered another pair for the 2nd build. This brings the Intel GPU part of the farm up to 6 machines.
Old specs
ASUS P6T motherboard
Kingston HyperX memory 12Gb, DDR3-1600 (3 x 4Gb sticks)
Intel core i7-970 (6 core/12 thread) 3.2Ghz
Corsair H80 water cooling system. Corsair fans replaced with Noctua fans
Intel 520 series 120Gb SSD
WD 500Gb HDD
Palit GTX660 graphics 2Gb, OC edition
Seasonic 750w power supply 80+ bronze
CM Storm Sniper case
New specs
ASUS X99-A motherboard
Kingston HyperX memory 16Gb DDR4-2133 (4 x 4Gb sticks)
Intel core i7-5820k (6 core/12 thread) 3.3Ghz
Corsair H80 water cooling system, Corsair fans replaced with Noctua fans
Intel 520 series 120Gb SSD
WD 1Tb HDD
2 x EVGA GTX750Ti graphics
Seasonic 750w power supply 80+ bronze
Fractal Designs ARC midi R2 case
As you can see almost the same, except its faster and has dual GPU's. The old machines will go up on eBay as a complete system unless I get an offer for them in the next fortnight or so.
Update 27th of October
I have dropped off the bits that I had to contribute to this 1st build. unfortunately its delayed for 2 weeks because memory is out of stock and got back-ordered.
25 October 2014
18 October 2014
18th of October
Farm news
I have finally reached 25 million credits for Seti so I am now scaling back and will concentrate on some of the other projects. Both Asteroids and ClimatePrediction need to catch up. In the case of ClimatePrediction they tend to go from feast to famine (ie lots of work to none) quite regularly. Asteroids have a CUDA app but because it does floating point maths it runs rather slowly. Both AMD and Nvidia limit the floating point capability of their consumer grade cards.
The farm is running during the day but the main GPU crunchers are usually off. Now that the dual Nvidia-GPU machine has been updated to GTX750Ti cards it may get to run during the day. I have already listed the two GTX660's on eBay.
Depending on how the 750Ti cards work out I may replace the remaining GTX660 cards. I have one GTX660 in each of the 6 core/12 thread crunchers.
Also on the shopping list are GTX970's to replace the GTX670's that are in the GPUgrid crunchers. I looked at them last week but decided they were a bit too expensive at the moment.
Project news - GPUgrid
They've decided to move to a CUDA 6.5 app and are dropping their 4.2 app. This fits with Nvidia ceasing support in their drivers for pre-Fermi cards. Given they are up to Maxwell architecture I expect Nvidia will drop support for the Fermi-based cards fairly soon.
I have finally reached 25 million credits for Seti so I am now scaling back and will concentrate on some of the other projects. Both Asteroids and ClimatePrediction need to catch up. In the case of ClimatePrediction they tend to go from feast to famine (ie lots of work to none) quite regularly. Asteroids have a CUDA app but because it does floating point maths it runs rather slowly. Both AMD and Nvidia limit the floating point capability of their consumer grade cards.
The farm is running during the day but the main GPU crunchers are usually off. Now that the dual Nvidia-GPU machine has been updated to GTX750Ti cards it may get to run during the day. I have already listed the two GTX660's on eBay.
Depending on how the 750Ti cards work out I may replace the remaining GTX660 cards. I have one GTX660 in each of the 6 core/12 thread crunchers.
Also on the shopping list are GTX970's to replace the GTX670's that are in the GPUgrid crunchers. I looked at them last week but decided they were a bit too expensive at the moment.
Project news - GPUgrid
They've decided to move to a CUDA 6.5 app and are dropping their 4.2 app. This fits with Nvidia ceasing support in their drivers for pre-Fermi cards. Given they are up to Maxwell architecture I expect Nvidia will drop support for the Fermi-based cards fairly soon.
17 October 2014
GTX750Ti upgrade
This is what we start with, a pair of GTX660's
And this is the GTX750's in their place.
Here are some GPU-Z screen shots. The first showing the card details
And this one showing it rather maxxed-out while running two Seti@home Multi-Beam cuda 50 work units.
They are certainly running a lot cooler than the GTX660's that were in there before. They also appear to be quicker but it could be the work units which vary in the amount of processing required.
12 October 2014
12th of October
Farm news
Not much happening on the farm during the day. Its usually too hot so the only things running are the Parallella's and Raspberry Pi's. At night I have been running the Intel-GPU machines. The Parallella and Pi's are running Einstein BRP4 work and the Intel's are still running Seti work.
On the hardware front I am going to get a pair of GTX750Ti's to replace some of the GTX660's. I will also be looking at replacing GTX670's which are in the GPUgrid crunchers. Given I have boxes of old graphics cards I will be listing these on eBay fairly quickly after they have been removed from the machines.
I got a rough price from the computer shop for the 20 thread Xeon servers, they were almost $3000 just for a CPU and then there was the motherboard. Needless to say I won't be buying one of them any time soon. I also looked at an i7-5820k (6 core/12 threads). The price was reasonable. But, and there is always a but, it uses 10 watts more than the i7-970's that its supposed to be replacing and is around the same clock speed. Sure it would be faster and has newer instructions.
Project news - Seti
They've released a new Astropulse application. It does a bit more maths and of course isn't compatible with the previous version. You'll either need to run the "stock" project supplied application or download the optimised versions and manually updated your app_info files.
There is an installer coming but will be a couple of days before that gets tested and released. In the mean time I have already installed it on the Intel GPU machines and managed to trash a bunch of multi-beam work units.
The estimated time to completion is way off while the server works out how fast your computer(s) are and will adjust itself once it gets 11 validated work units, excluding the ones that are more than 10% blanked or otherwise fail.
Not much happening on the farm during the day. Its usually too hot so the only things running are the Parallella's and Raspberry Pi's. At night I have been running the Intel-GPU machines. The Parallella and Pi's are running Einstein BRP4 work and the Intel's are still running Seti work.
On the hardware front I am going to get a pair of GTX750Ti's to replace some of the GTX660's. I will also be looking at replacing GTX670's which are in the GPUgrid crunchers. Given I have boxes of old graphics cards I will be listing these on eBay fairly quickly after they have been removed from the machines.
I got a rough price from the computer shop for the 20 thread Xeon servers, they were almost $3000 just for a CPU and then there was the motherboard. Needless to say I won't be buying one of them any time soon. I also looked at an i7-5820k (6 core/12 threads). The price was reasonable. But, and there is always a but, it uses 10 watts more than the i7-970's that its supposed to be replacing and is around the same clock speed. Sure it would be faster and has newer instructions.
Project news - Seti
They've released a new Astropulse application. It does a bit more maths and of course isn't compatible with the previous version. You'll either need to run the "stock" project supplied application or download the optimised versions and manually updated your app_info files.
There is an installer coming but will be a couple of days before that gets tested and released. In the mean time I have already installed it on the Intel GPU machines and managed to trash a bunch of multi-beam work units.
The estimated time to completion is way off while the server works out how fast your computer(s) are and will adjust itself once it gets 11 validated work units, excluding the ones that are more than 10% blanked or otherwise fail.
Subscribe to:
Posts (Atom)