HCC GPU Launched.

[Conquistador SW]

I found a post on the WCG boards reporting about 4 min/wu for a 7750.
https://secure.worldcommunitygrid.or...ad?post=395655

Thanks!

If that is true, that would also make the 7750 about 2 times slower in which case the 7770 has the edge in points/watts. Looking at the specs I was thinking it would do better.

Edit: according to another post he made later on in that thread the GPU takes 3 minutes and his athon x4 adds another minute, so maybe with a faster CPU you can do these units in less than 4 minutes with the 7750, but I guess the 7770 will still be a better pick.

[Rob_B]

I got another kill-a-watt like device to check the power usage of the GPU.

I was running this rig without CPU work and 2 GPU WU's at a time. I was using 205 watts with 2 WU's on the GPU. And with one WU's suspended it was using 193 watts.

Now I have switched back to do CPU work and 1 GPU WU at a time. With the CPU WU's suspended (only GPU work being done) it uses 186 watts and with the GPU idle it drops to 131 watts.

The power consumption between 1 or 2 WU's is just 20 watts and between 1 WU and idle is 55 watts.
When running 2 GPU WU's at a time it takes 4:40-5:00 to complete a WU. And with 1 WU at a time it takes 3:00-3:15 to complete a WU.

This is on a i7-860 @ 4 GHz (no HT) with a ATI 5850 @900/1000 MHz.

[Mydog]

Now food for thought, what would an SR-X with two 15 core IB Xeons paird with four 7990 dual GPU cards do in WCG and Boinc? One mill WCG PPD?

[masterg]

I have a

q9550 + 5850

3770k + 5870

e3 1235

this normally nets me 12k boinc a day.

I threw down 32k yesterday >: )

and just added another 5830 to my e3 1235

[Duh]

will an ati 6670 with HF acetal extreme should it have any problem to run 24/7 ? I've had 2 or 3 gpu cards die before

[OldChap]

If you can keep the vrm's cool I think you are good.

A couple of hours of playing around with these cards and I get the following info:

As seen before the 5870 on q6600 @ 3.3 gives a runtime of somewhere around 3m 20s. the cpu time at end of wu is maybe 53s plus

running on 3770K @ 4.8 the average time for 20 wu's was 2m 57s and the second cpu time is around 41s

Conclusion: GPU will breath new life into an old rig and make it useful for crunching but if you can use a newer, faster rig you will definitely see benefits. My spreadsheet suggests nearly 2k ppd between these two.

Power use:

MVG, 3770K at 4.8, 8G samsung green @ 2000, Samsung spinpoint hj103 plus 1 fan running together with an asus hd5870 Idles in windows at 100W

Running 8 x GFAM takes this to 210 watts and would net 10K+ ppd in linux.

Running 1 x GPU plus 7 x GFAM takes the power use to 270 watts where I predict it will produce around 15K extra ppd

Running 2 x GPU plus 6 x GFAM takes power use to 360 watts so a single card would seem to use 90 watts to produce 15K. This figure takes into account the losses expected by running less cpu tasks

This setup is running on a 650w seasonic x ..... 55% of rated power now which is right in the sweet spot

Conclusion: this single rig with simple old cards running mildly overclocked on cores to 900 seems likely to produce in the order of 32K GPU + 7K5 CPU.

My new processors arrived today... the new board I have my eye on has 7 slots. If the 7770's are about as power hungry as these cards I could end up with all my crunching on one machine that does 125K ppd with current cards, 160Kppd with mid range cards or 230K ppd with top end cards

EDIT:Tomorrow I will investigate the implications of running HT off.... With 2 cards running it may be break even but with more ...we will see

[retro77]

I wonder how long the WUs will last?

[Mydog]

I wonder how long the WUs will last?

I hope they last a long time but with the speed GPU's are chewing through them they might not.

[0ne.shot]

Okay, so I did the math. If I run my GTX 570 with a 400W load, or a 7770 at 200W load, it'll take (6000 hours)250 days of 24/7 constant running to equal the cost of running a GTX 570. The cost is factored in with $0.10/KWh. However, the 7770 will also produce many more points/KWh, too, so that has to be factored in.

[shoota]

I see a lot 5000 and 7000 series ATI cards but where are the numbers for the 6000 series?

[OldChap]

Are your calculations based on measured load running HCC GPU?

Manufacturers ratings are no good here just now .... I think mine are rated to pull around 200w here it is only around 90w

With power use remember too that a proportion of the wu runtime is not with GPU under any load so again that is less cost.


I've yet to see any power use figures for 6000's but looking at jspace's results the 6970 should fall near the 7770 for performance

[werdwerdus]

IMO the workunits need to be about 10x longer (more work done) than they current are. For comparison the GPUGrid tasks take 6-12 hours on decent hardware.

[Movieman]

IMO the workunits need to be about 10x longer (more work done) than they current are. For comparison the GPUGrid tasks take 6-12 hours on decent hardware.

I tend to agree with you but not to that degree..10 mins would be fine BUT they did this thinking about heat and that is understandable also.
They think worst case, IE: Mrs Jones Dell that hasn't been cleaned inside since new..

[RaynQuist]

My 6870 takes around 2:30 per WU, but my stock i7 920 adds ~30 seconds in the beginning and ~1:30 at the end

[retro77]

Too bad the IGPU in the 2600k is not supported....

[0ne.shot]

Are your calculations based on measured load running HCC GPU?

Manufacturers ratings are no good here just now .... I think mine are rated to pull around 200w here it is only around 90w

With power use remember too that a proportion of the wu runtime is not with GPU under any load so again that is less cost.


I've yet to see any power use figures for 6000's but looking at jspace's results the 6970 should fall near the 7770 for performance

I was making an estimation based off of this: http://www.anandtech.com/bench/Product/536?vs=518

At the bottom it shows load temps with OCCT on both GPUs. A 2700k @ 4.7GHz and GTX 570 load is right at 400W - sometimes just under or over. The 7770 looks to be right around 200W. I'm guessing HCC GPU isn't as stressful as OCCT, but the power consumption numbers still should be relative to the HCC GPU tasks. I have an APC UPS that tells me current power draw. Of course, that is from the wall and I'm not certain of the PSU Efficiency @ 400W with a Corsair TX750.

I have no idea what the manufacturer's power rating are.

I just finished a HCC GPU WU on my A-3650 APU with Radeon 6530D (320SPs) @ a total of 13:27. It takes quite awhile and the UPS on that PC says 90W with a HCC GPU task and 3 HCMD 2 tasks running.

Too bad the IGPU in the 2600k is not supported....

It is a bummer. The iGPU on my A6 3650 works but not on this 2700k.

[kjeldoran]

Taken from the WCG forum.

<app_info>
<app>
 <name>hcc1</name>
 <user_friendly_name>Help Conquer Cancer</user_friendly_name>
</app>
<file_info>
 <name>wcg_hcc1_img_6.56_windows_intelx86__ati_hcc1</name>
 <executable/>
</file_info>
<file_info>
 <name>hcckernel.cl.6.56</name>
</file_info>
<app_version>
 <app_name>hcc1</app_name>
 <version_num>656</version_num>
 <platform>windows_intelx86</platform>
 <avg_ncpus>1.000000</avg_ncpus>
 <max_ncpus>1.000000</max_ncpus>
 <flops>32282641180.000000</flops>
 <plan_class>ati_hcc1</plan_class>
 <api_version>7.1.0</api_version>
 <file_ref>
  <file_name>wcg_hcc1_img_6.56_windows_intelx86__ati_hcc1</file_name>
  <main_program/>
 </file_ref>
 <file_ref>
  <file_name>hcckernel.cl.6.56</file_name>
  <open_name>hcckernel.cl</open_name>
 </file_ref>
 <coproc>
  <type>ATI</type>
  <count>0.500000</count>
 </coproc>
 <gpu_ram>262144000.000000</gpu_ram>
</app_version>
</app_info>

Use at your own risk. app_info only usable for HCC GPU only.

It works for my 5850, but now it only uses the GPU. No work for the CPU. And it went up from 3 minutes per WU to over 5 minutes.

I started to use this app_info this evening. Yes, you lose your work for your CPU, but if you are focusing only on HCC then you will complete more WUs than using both your CPU and GPU based on my observations with my setup.

Previously, I was doing GPU WUs around 1:56 with only one WU and 7 CPU threads. Using the above app_info, I lose all of my CPU threads but I work on 2 GPU WUs at once. Completion times around around 2:06 to 2:20 per WU. Yes, it is slower but I am completing two WUs at once. Even if I had the max of 2:20 per WU that is over 1200 WUs in one day. Of course, there will be variations and even if it was only 1100 WUs it is still more than what I would get with 7 CPU threads and 1 GPU thread running HCC (which is around 900 WUs in one day at 1:56 per GPU WU). I'd really only recommend the above app_info if you run HCC only and your GPU is able to handle the work.

You can also do more than 2 GPU WUs at once if you change <count>0.500000</count> to .33 (3 at once) or .25 (4 at once). Info was taken from nanoprobe on WCG forums for increasing the number of WUs beyond 2.

Pros:
Decreased power consumption since only 2 threads are being used on the CPU
More WUs completed

Cons:
Only able to run HCC GPU at this point in time, maybe there can be code added to run another project later on.
More stress on the video card
Some video cards may not benefit from this app_info

I am still in the testing phase of this and I will see how the results are in the next few days. I might see if changing the count to 3 or 4 can increase overall WU completion.

Last thing, I would imagine using this tweak will show a greater effect when the core clock is increased since the GPU is being utilized more.

[OldChap]

Thanks!

If that is true, that would also make the 7750 about 2 times slower in which case the 7770 has the edge in points/watts. Looking at the specs I was thinking it would do better.

Edit: according to another post he made later on in that thread the GPU takes 3 minutes and his athon x4 adds another minute, so maybe with a faster CPU you can do these units in less than 4 minutes with the 7750, but I guess the 7770 will still be a better pick.

Gentlemen. Would you all be kind enough to publish your CPU times. I have a feeling that we will see that regardless of GPU used we can get a clear picture of the effect of platform. So, what is needed is from observation ...

primary cpu time, secondary cpu time, cpu type, cpu speed and whether ht is on or off.

EDIT: HT results




Given that with ht on there are 4 threads of cpu running then there is no good reason to move to ht off with a single card of this type, however when running 2 cards the advantage seems to be in running ht off

I suspect this may also have an effect on the decision to run multiple instances on certain cards.

[mreuter80]

...
I just finished a HCC GPU WU on my A-3650 APU with Radeon 6530D (320SPs) @ a total of 13:27. It takes quite awhile and the UPS on that PC says 90W with a HCC GPU task and 3 HCMD 2 tasks running.

Quick question: is that with the iGPU only or in combination with a discrete card?
Would love to see more APU numbers, since I'm researching currently in this direction for a new HTPC build.

[jspace]

Does anyone know how to read exact times out of the results? I'm guessing it's in the tail end.

cl_khr_d3d10_sharing
Estimated kernel execution time = 0.31736 [sec]
Starting analysis of ../../projects/www.worldcommunitygrid.org/X0960069160298200605241517_X0960069160298200605241 517.jp2...
Extracting GLCM features...
Total kernel time: 98.204834 (1026 kernel executions)
Total memory transfer time: 3.162154
Average kernel time: 0.095716
Min kernel time: 0.087317 (dx=24 dy=7 sample_dist=24 )
Max kernel time: 0.104350 dx=2 dy=2 sample_dist=2
INFO: GPU calculations complete.
Total time for ../../projects/www.worldcommunitygrid.org/X0960069160298200605241517_X0960069160298200605241 517.jp2: 197 seconds
Finished Image #0, pctComplete = 1.000000
CPU time used = 88.998571
08:32:57 (4316): called boinc_finish

</stderr_txt>
]]>

Edit: Let's add a few more results. I only have one machine capable of running theses atm. No power consumption figures.

cl_khr_d3d10_sharing
Estimated kernel execution time = 0.33044 [sec]
Starting analysis of ../../projects/www.worldcommunitygrid.org/X0960069161475200605171506_X0960069161475200605171 506.jp2...
Extracting GLCM features...
Total kernel time: 116.657501 (1026 kernel executions)
Total memory transfer time: 1.940834
Average kernel time: 0.113701
Min kernel time: 0.104511 (dx=23 dy=11 sample_dist=24 )
Max kernel time: 0.122929 dx=2 dy=0 sample_dist=1
INFO: GPU calculations complete.
Total time for ../../projects/www.worldcommunitygrid.org/X0960069161475200605171506_X0960069161475200605171 506.jp2: 213 seconds
Finished Image #0, pctComplete = 1.000000
CPU time used = 92.024990
08:22:55 (12696): called boinc_finish

</stderr_txt>
]]>

cl_khr_d3d10_sharing
Estimated kernel execution time = 0.32184 [sec]
Starting analysis of ../../projects/www.worldcommunitygrid.org/X0930069660332200606081307_X0930069660332200606081 307.jp2...
Extracting GLCM features...
Total kernel time: 92.362144 (1026 kernel executions)
Total memory transfer time: 1.823963
Average kernel time: 0.090022
Min kernel time: 0.082192 (dx=25 dy=4 sample_dist=24 )
Max kernel time: 0.098286 dx=2 dy=2 sample_dist=2
INFO: GPU calculations complete.
Total time for ../../projects/www.worldcommunitygrid.org/X0930069660332200606081307_X0930069660332200606081 307.jp2: 188 seconds
Finished Image #0, pctComplete = 1.000000
CPU time used = 91.135784
05:46:52 (12684): called boinc_finish

</stderr_txt>
]]>

[Plan.B]

I'm finishing a wu about every 2.5 minutes w/ gtx580 sli + 2600k. All clocks are stock. 6 cpu threads (any project) & 2 gpu threads + 2 cpu threads (HCC gpu).

[bluestang]

With a KILL A WATT....my Home PC setup in sig is drawing 186W full load CPU only (running GFAM and HCC GPU).
When GPU kicks in for HCC GPU its at 249W
And running with 22" display and 2.1 speakers I'm at 289W.

2600k @ 4.5GHz, 7870 @ 1050/1250

I'm running anywhere from 1:31 to 1:53 with most of them in the 1:3x range. On the E7500 machine it was 1:32 to 1:57 with most in the 1:5x range. So an improvement using the 2600k.

When do a common 1:32 WU...

15 sec CPU
37 sec GPU
40 sec CPU

Now, I can get them in the 1.2x range if I clock my 7870 at 1200/1400. I let in run like that overnight and was amazed. However, I started using my pc this morning and after a few minutes a solid grey screen appeared and locked that way. Had to reset to restart PC. May have a bum card. Now running at 1050/1250 stk -- not very Xtreme.

[bluestang]

I want to try 2 GPU WU at once. How do I accomplish that and still run 6 WU GFAM on CPU?

[0ne.shot]

Quick question: is that with the iGPU only or in combination with a discrete card?
Would love to see more APU numbers, since I'm researching currently in this direction for a new HTPC build.

That is with the iGPU only. The APU is not overclocked and stock CPU Clock speed is 2600 MHz. Youtube videos lag badly when running HCC GPU and make the PC unusable. I'm going to check the CPU/GPU/CPU times on the project and I'll post them up shortly.

Update: I watched a few WUs at the start and toward the end, and it looks like the first ~40 seconds and last ~60 seconds are CPU time and the rest is GPU. This was observed on 2 WUs that completed in ~12:40 seconds. It's hard to gauge when the CPU started and GPU stopped as the CPU utilization barely changes through the entire run time of the WU.

[OldChap]

bluestang: I believe that it is not possible to run 2 wu at once and have cpu wu's running too