Could be interesting to see the times with different cards and settings.

HD 5850 @810 MHZ.

First WU finnished in 6000 seconds.
second is running for 2 hours now and its only at 56%

---

With each crime and every kindness we birth our future.

Could be interesting to see the times with different cards and settings.

HD 5850 @810 MHZ.

First WU finnished in 6000 seconds.
second is running for 2 hours now and its only at 56%

It strongly depends from blanking % value. Senseless to post any time w/o saying how much blanking this particular task has. Think about it just as AR value for MB task.

4670 (stock) using OpenCL r456

0% blanking ~26,000 seconds
~4% blanking, ~26,400 seconds

Compared to i5 750 (stock) using r406

0% blanking, (4x) ~34,500 seconds

She isn't breaking any world records but, its considerably faster than a single core on my i5.

---

I've got a 5850 and from memory 0% blanking take about 2 hours
15% takes about 4 hours and 25-30% takes around 6 hours

---

In a rich man's house there is no place to spit but his face.
Diogenes Of Sinope

The new 6970 is king of the hill (imho).

You can run several work units in parallel. At the moment I am running 4 work units in parallel. One unit (zero blanking) finishes in about 8.000 seconds.

That's 2.000 seconds per unit (or 33 minutes!).

http://setiathome.berkeley.edu/results.php?hostid=5694841&offset=0&show_names=0&state=2&appid=5

Running multiple instances in parallel also mitigates the cost of blanking - and memory latency - to some degree.

the 6XXX series is a bit slower than the 5XXX series unfortunately. I'm not sure why they would hobble the latest GPU's but that is what appear to have happened at ATI.

---

In a rich man's house there is no place to spit but his face.
Diogenes Of Sinope

I've got a 5850 and from memory 0% blanking take about 2 hours
15% takes about 4 hours and 25-30% takes around 6 hours

I just finished a heavily blanked AP WU

there appears to be a bit of improvement in the latest Lunatics app version. This WU was blanked over 70%

---

In a rich man's house there is no place to spit but his face.
Diogenes Of Sinope

the 6XXX series is a bit slower than the 5XXX series unfortunately.

?

Don't get confused.

They (AMD marketing) simply changed the naming scheme. E.g the successor of 5870 is not 6870 - it's 6970 :

5850 -> 6950
5870 -> 6970
5970 -> 6990 (Q1/2011)

The new cards are faster. But a little bit more power hungry. But also better performance per wattage.

Few times I've seen on one of my 5870s running at 900/1150:
11,748.26 blanking 82.22
4,830.98 blanking 2.34
2,786.47 blanking 0

-Dave

---

the 6XXX series is a bit slower than the 5XXX series unfortunately.

?

Don't get confused.

They (AMD marketing) simply changed the naming scheme. E.g the successor of 5870 is not 6870 - it's 6970 :

5850 -> 6950
5870 -> 6970
5970 -> 6990 (Q1/2011)

The new cards are faster. But a little bit more power hungry. But also better performance per wattage.

Not the same GPU Note the shaders. The 6870 is a significantly slower GPU compared to the 5850. All you need to do is check on the pricing. Hint: I have. The 6XXX series are all significantly cheaper than the 5850,5870 and 5970. So you have a brand new GPU yet its cheaper than the older GPU's hmmm I'd say they are stepping back on their GPU's

---

In a rich man's house there is no place to spit but his face.
Diogenes Of Sinope

skildude, I don't see your point.
Frizz can run AP task per 33 mins on HD6970. Can you do the same ot better with HD5870?
Cheaper GPU sometimes means more agressive pricing policy, not performance degradation...

Skildude, you're mixing up the 6800 and 6900 series...
6800 = slower than 5800
6900 = faster than 5800

See here for example.

-Dave

---

Not the same GPU. Note the shaders. The 6870 is a significantly slower GPU compared to the 5850.

???????

Yep - and because of this the 5850 should/must/can be compared to the 6950. 9. NINE. Not to the 6850 or 6870.

Scroll 5 lines up: http://setiathome.berkeley.edu/forum_thread.php?id=62385&nowrap=true#1057104

Besides:
57xx / 68xx = no DP
58xx / 69xx = DP

All this confusion just because AMD (marketing) decided to - sort of - unify it's naming scheme.

I have done some extensive testing. Here are my results.

- Values are in seconds per 0.9% progress (multiply by 111 and you get the final runtimes).

- Values in brackets are normalized runtimes in case more than one tasks runs in parallel.


What you can see is, for example,

- when you compare GTX570 to HD6970: Initially the GTX570 is faster - but when running more tasks in parallel the HD6970 scales better.

- for the HD6850 it makes no sense to run more than 3 units in parallel. HD6870 scalles a little bit better.

- HD6970 scales really well. Will try 5 or more tasks later (shouldn't be a problem with it's 2GB of memory).


What you can't see is (I've done some more extensive testing, but that's to much stuff to post it here):

- for lower AMD cards it's favorable to run CPU tasks in addition (MB, AP). I got my best RAC with 3 x AP GPU plus 4 x AP CPU task.

- for GTX570 and HD6970 it's better to run no CPU work. Just keep feeding the GPU. Otherwise the GPU runtimes increase a lot - and overall performance decreases. At least with my CPU (AMD Phenom II X4 965).


GTX570:

1 x GPU: 26

2 x GPU: 47 (23,5)

3 x GPU: 67 (22,333)

4 x GPU: unstable

---------------------

HD6850:

1 x GPU: 38

2 x GPU: 63 (31,5)

3 x GPU: 87 (29)

4 x GPU: 116 (29)

---------------------

HD6870:

1 x GPU: 35

2 x GPU: 57 (28,5)

3 x GPU: 82 (27,3)

4 x GPU: 108 (27)

---------------------

HD6970:

1 x GPU: 28

2 x GPU: 46 (23)

3 x GPU: 64 (21)

4 x GPU: 72 (18)

---
All tests done with Windows XP32 (because it's about 10% faster then, for example, Windows 7 x64).

But 6970 / newest AMD Catalyst doesn't run stable under Win XP32 / OpenCL (driver hangs, blue screens, reboots, ... real ugly stuff). So I had to switch to Windows 7 x64 for production mode. And - SURPRISE - 6970 runs faster under Windows 7 x64 ... really strange. Only AMD knows why ;-)

Frizz - do you do any other stuff on your 6970 system when it's doing AP tasks? Or is it "crunching only"? If not, have you noticed any screen lag or delays?
(I occasionally do with my HD 5870 that's on my machine that I use all the time).

-Dave

---

3 x GPU: 87 (29)

4 x GPU: 87 (29)

how it can be? either elapsed or normalized is wrong here.

3 x GPU: 87 (29)

4 x GPU: 87 (29)

how it can be? either elapsed or normalized is wrong here.

You're right. Corrected it.

Frizz - do you do any other stuff on your 6970 system when it's doing AP tasks? Or is it "crunching only"? If not, have you noticed any screen lag or delays?
(I occasionally do with my HD 5870 that's on my machine that I use all the time).
-Dave

Hello Dave,

it's a dedicated cruncher - so I don't care about lags :)

When running only one task on the 6970 there are almost no lags. When you run 4 in parallel you can't really work any more on that PC.

Because of your problems: have a read here:
http://lunatics.kwsn.net/12-gpu-crunching/astropulse-for-ati-gpus-released.0.html
...
-ffa_block 8192 (default value) - defines how many different periods GPU will process per single kernel call
-ffa_block_fetch 2048 (default value) - defines how many threads will be used in FFA initial fetch kernel
Rules for using these values:
-ffa_block_fetch <number> can be used only if -ffa_block <number> already listed in command line
numbers should be even,better if they will be power of 2, ffa_block should be divisible by ffa_block_fetch.
If you experience lags during application execution try to decrease these values.
...

Frizz - do you do any other stuff on your 6970 system when it's doing AP tasks? Or is it "crunching only"? If not, have you noticed any screen lag or delays?
(I occasionally do with my HD 5870 that's on my machine that I use all the time).
-Dave

Hello Dave,

it's a dedicated cruncher - so I don't care about lags :)

When running only one task on the 6970 there are almost no lags. When you run 4 in parallel you can't really work any more on that PC.

Because of your problems: have a read here:
http://lunatics.kwsn.net/12-gpu-crunching/astropulse-for-ati-gpus-released.0.html
...
-ffa_block 8192 (default value) - defines how many different periods GPU will process per single kernel call
-ffa_block_fetch 2048 (default value) - defines how many threads will be used in FFA initial fetch kernel
Rules for using these values:
-ffa_block_fetch <number> can be used only if -ffa_block <number> already listed in command line
numbers should be even,better if they will be power of 2, ffa_block should be divisible by ffa_block_fetch.
If you experience lags during application execution try to decrease these values.
...

Cool, thanks for the tip! It's not really bad, but I may decide to make a change, will see...

-Dave

---

skildude, I don't see your point.
Frizz can run AP task per 33 mins on HD6970. Can you do the same ot better with HD5870?
Cheaper GPU sometimes means more agressive pricing policy, not performance degradation...

First I don't have a 5870 Sooo no I can't second. A next gen GPU which starts lower and doesnt really stomp the previous gen. Really thats a vast difference. These could have easily been labeled as 5851, 5871 etc. since they clearly aren't using more shaders and the GPU speed is essentially unchanged. The only real change is the Memory speed. 5850 =4000Mhz 6870 = 4200mhz. also note there are 320 fewer streaming processes on the 6870. It seems like these GPU's are just a lateral shuffle of ability and are just a lame excuse to relabel some GPU's Notice that the numbers are 6XXX I'd expect a boost somewhere. Perhaps its in the architecture. So far I've not seen any compelling argument to make me think otherwise

---

In a rich man's house there is no place to spit but his face.
Diogenes Of Sinope