@skgiven

Speaking of GPU crunching , I know that you ran MilkyWay and GPUGrid
and I wonder if you were using ATI or Nvidia , and if they were aircooled
or watercooled ?
I haven't done MW for a while but when I was , I was mostly running my
GTX 285 and a little on a GTX 295. But The 295 dual GPU just got WAY too hot
so I stopped using it. That is why I wonder if yours are aircooled.
I have since then replaced the 285 with a GTX 560 Ti Classified 448 Core and I wonder
if it should be OK on Air ? Also I wonder if anyone is using the 560 over at MW
and how it is performing with the DP ?

Think that's what he meant to say, but 1 Card, 1 Day max, as a machine can have 2,3,4,6,8 GPUs and even a mix of brands in them [a tweakers heaven to get that to crunch on all cylinders].

Still, if the card allowed to be subdivided into multiple logic processing units and do 2 or more simultaneous, I'd not have trouble with clocking 2 times [is a HT capable CPU not doing the same to acknowledged runtime?] At the end of the day, having found how I can tell my NVidia card to stay at maximum performance, a serious blazing GUI experience generating right from the first nanosecond, it's the number of results returned we're all after. In BOINC world only cobblestones count, WCG will surely do it's thing to have the runtime tally be equitable, and maybe set new standards how to reasonably compensate in inconvertibles. We'll work through whatever comes up to get an amenable solution. Silently, the techs surely will pick out the valid and reasonable points from the discussions and do the testing in the background in their labs. :D

--//--

tomast, I stick to air cooling, but it’s as much a location thing as optimizing the hardware. Air is cheaper and less complicated, but you do need to clean the GPU fairly regularly. I have used both NVidia and AMD/ATI at MW but ATI’s superior DP makes it king there. I have also used both for Folding, but the performance of GPU’s is poor there compared to CPU’s.

The GTX 560 Ti Classified 448 Core is an excellent card. Due to DP insufficiency Fermi's do not perform very well at MW, but it would still be a reasonable performer there. In SP (including here) it will perform much better. If it has a dual fan your 448 is probably going to run reasonably cool, but I always use fan controlling software, where possible, and I recommend everyone does this on high end cards. For Windows you can use NVidia System Tools, MSI Afterburner, EVGA Precision...

PS. Each time you restart a Linux system you need to set the GPU to prefer maximum performance, from NVidia control panel, and I think Coolbits still doesn't work for Fermi's.

First of all, thank you all for discussing in this thread the various tradeoffs in different approaches. There are flaws in all of our measurements (points, runtime, results) as they can all reflect different things rather than just relative contribution.

We are focused on making GPU work for our volunteers. This is priority number one. I had originally hoped to tackle other items on our to-do list with the work as well (such as making 1 WCG point = 1 BOINC credit that is still leftover from when we shutdown UD). However, we are putting a stake in the ground to update the server code and shortly thereafter start beta testing GPU. As a result instead of trying to solve everything, we are only doing the essential so that we can get it running.

An example of this is that the new credit logic from BOINC appears to work quite well for consistently granting credit across all platforms (in alpha the ATI and Nividia GPU versions are claiming within 5% of Linux CPU which is claiming within 5% of Windows ) provided the flops estimates are consistent (if they aren't, then things behave poorly). There is a short 'ramp up' phase were baseline stats are collected where things are a little out of alignment, but that is short lived.

So we are simply going to jump into using the new credit logic. There is going to be a disconnect from credit before the server code update and credit after the server code update. Our testing suggests it will be pretty close, but we will watch it after launch. If a more credit is awarded, then we will probably leave it (this is a request from David Anderson to be consistent with the overall BOINC community). If it is less, we will probably tweak it up some to be consistent with the old credit we awarded.

Once we have stabilized things after the server code update and we see how things are doing in beta testing we will look at 'balance' issues. As an early preview on our thoughts, credit/points has always been the metric most designed to reflect the contributed computing power directly (and thus reflects whether you are contributing time from PII or a i7) while runtime reflects the time that a workunit was allowed to be running on a resource. Any policy changes that we make need to not only consider cpu vs gpu, but also consider the impact of contribution from Android on ARM and computers with cpus that have cores that perform at different speeds that will be part of the environment over the next 18-24 months. Additionally, given the explosion of cores on machines, we are starting to look at if any of the apps we have run or will be running can be multi-threaded and use more than one cpu cores (note - a lot of the apps we have are not designed to be multi-threaded so this can only be done if the app itself is already multi-threaded). This will become important to reduce the RAM being used on volunteers machines with many concurrently running tasks. We expect to have a range computers that have a 100-1000 fold performance difference between them over the next 24 months. There are going to be a lot of challenges determining what is 'fair' and yet be motivational for all contributors over that period.

Thanks for the update. Good luck with the server and credit changes. Good to know you are pondering the performance impact of such future devices, trying to facilitate greater variety and systems with relatively limited memory.

tomast, I would expect you will be able to use the GPU for HCC and select to not use the CPU for HCC; on Boinc projects that allow you to select to use the GPU, or not, you can also select to not use the CPU (MW for example). Also, HCC might not even be available for CPU's when the GPU project starts (unless the same app can run on the CPU) and even if it was you can presently deselect HCC for CPU.
Cheers for the coolbits 4 feedback ;)