Message boards :
Number crunching :
ATI Radeon HD5970 performance on boinc
Message board moderation
Previous · 1 · 2 · 3 · 4 · Next
Author | Message |
---|---|
Send message Joined: 8 Nov 07 Posts: 323 Credit: 1,362,120 RAC: 0 |
|
Send message Joined: 14 Dec 09 Posts: 161 Credit: 589,318,064 RAC: 0 |
OC: Core 850, Memory 1200 and the stock cooler is insufficient at full capacity if running 24/7, temperature is around 93 celcius and going further with oc, the workunits take longer. I suggest the values are upper limit with the air cooler. |
Send message Joined: 12 Aug 09 Posts: 172 Credit: 645,240,165 RAC: 0 |
Well I have my twin 5970's running flat out on MW now. I got around the three GPU problem by enabling Crossfire AND connecting the Crossfire cable. The problem is that I cannot run any Collatz as the system slows to like 45 hours per WU. Running one at a time I get times of around 6:52 per WU. Running two increases the time to ~14 minutes per WU. Three runs out to ages and a really jerky mouse. I expect I could restore Collatz by simply removing the bridge, but sheesh, talk about manual control lol. I expect my oil cooling blocks to arrive at the end of this month, so it will be interesting to see how these go. In the meantime I am swapping out all the Nvida 295, 260 crap and replacing it with another couple of 5970's, so I can reduce my power consumption by a coupe of hundred watts and increase throughput. It's all good. hahaha. |
Send message Joined: 30 Dec 07 Posts: 311 Credit: 149,490,184 RAC: 0 |
OC: Core 850, Memory 1200 and the stock cooler is insufficient at full capacity if running 24/7, temperature is around 93 celcius and going further with oc, the workunits take longer. I suggest the values are upper limit with the air cooler. No surprise that it's hot if you are running MilkyWay ATI with memory 1200. There is minimal to no advantage in running MilkyWay ATI with a high memory speed. All you are doing is producing extra heat, using more power and perhaps even limiting your maximum possible OC on the core. For MilkyWay I currently use core 960 and memory 500 on a 5870. |
Send message Joined: 14 Dec 09 Posts: 161 Credit: 589,318,064 RAC: 0 |
OC: Core 850, Memory 1200 and the stock cooler is insufficient at full capacity if running 24/7, temperature is around 93 celcius and going further with oc, the workunits take longer. I suggest the values are upper limit with the air cooler. Are you suggesting to lower the memory and use the room for increasing the core clock? |
Send message Joined: 30 Dec 07 Posts: 311 Credit: 149,490,184 RAC: 0 |
Yes, try it for yourself and see. A lower memory speed should reduce the card's temperature when running MilkyWay. This has been my experience with a 4890 and a 5870 so it should also apply to a 5970. A reduced temperature may allow you to clock the core higher. The results will probably vary from card to card. Even if you are not able to exceed 850 core speed then you may still benefit from reduced temperature and lower power consumption. The speed of processing of MilkyWay ATI tasks is not affected very much at all by memory speed. This is not the case with Collatz Conjecture ATI tasks which require high memory bandwidth for the best performance. I use AMD GPU Clock Tool to set the core and memory speed because MSI Afterburner would not allow a reduction of the memory speed. |
Send message Joined: 14 Dec 09 Posts: 161 Credit: 589,318,064 RAC: 0 |
For sapphire 5970, the least memory clock speed is 1010. Although i lowered the memory while keeping the core clock as same as the initial oc value (850), the temperature didn't change. I believe i have to work more around that:). Thanks for the tip. |
Send message Joined: 9 Nov 08 Posts: 41 Credit: 92,786,635 RAC: 0 |
My HD5970 goes at 775Mhz Core and 300 Mhz memory - at default voltage and fan speed. A proud member of the Polish National Team COME VISIT US at Polish National Team FORUM |
Send message Joined: 30 Dec 07 Posts: 311 Credit: 149,490,184 RAC: 0 |
For sapphire 5970, the least memory clock speed is 1010. Although i lowered the memory while keeping the core clock as same as the initial oc value (850), the temperature didn't change. I believe i have to work more around that:). Thanks for the tip. Yes Catalyst Control Center and MSI Afterburner both have limitations when adjusting the memory speed. If you have used MSI Afterburner it may have changed the default range of adjustment for the memory slider in Catalyst Control Center preventing you from reducing the memory speed to a low level. This is why I use AMD GPU Clock Tool to adjust core and memory speed. If you are using only Catalyst Control Center then PowerPlay will still be working and the memory speed shown in the details of the task will be incorrect. It will show a low value of 300MHz or 500 Mhz memory speed or something similar depending on how the idle clocks have been set for that card model, even though the task was actually processed with memory running at 1000 or higher. You need to check the memory speed with GPU-Z while a task is being processed to see the actual memory speed that is being used. If you use AMD GPU Clock Tool, then PowerPlay is disabled so this is not an issue. The only limitation is you need to reset the core speed with AMD GPU Clock Tool every time you reboot. This is all based on my experience with a 5870, so some things may not apply or may have to be done differently for a 5970. As well as the obvious clock speed differences, to prevent possible damage 5970s have tighter regulation of current/temp levels than 5870s which can cause the memory speed to be throttled when the limits are exceeded. Therefore sometimes a high memory speed on a 5970 can give less performance on applications or benchmarks than the default memory speed because the memory speed has been throttled by the hardware overcurrent/over temp regulation. There are some compromises necessary when designing dual core cards, so chasing the very last bit of performance is not really worthwhile for some people and a more moderate level of overclocking will still yield excellent performance, with lower temperatures, greater stability and less power usage. |
Send message Joined: 14 Dec 09 Posts: 161 Credit: 589,318,064 RAC: 0 |
I have been experimenting with MSI Afterburner and with 1.162 core clock and default memory voltage, 925/915 is stable and each wu take 76.9~77.5 seconds. At this point, each time i lower the memory clock, the performance drops. For example, if i set 925/880, again significant performance drop. In previous posts, cyclo stated a 970/300 configuration on 5870. This is very interesting and I assume the way memory is dealt in 5970 is different than 5870. |
Send message Joined: 30 Dec 07 Posts: 311 Credit: 149,490,184 RAC: 0 |
That is interesting, thanks for reporting your findings. It is useful information for other 5970 owners. I wonder if it's because 2 cores are accessing the memory on a 5970 rather than 1 core on a 5870. Looks like what I was suggesting is not totally correct for 5970s. At least reducing the memory speed to 915 allowed you to increase the core clock to 925 with stability, so perhaps I was partly correct. 925 core with stability when processing MilkyWay tasks is an excellent overclock for a 5970 so you have done very well with your overclocking experiments. I just tried MSI Afterburner again and it now allows me to set 500 memory speed on my 5870 whereas before it didn't. Strange, must be due to the way default clock profiles get set by Catalyst Control Centre and AMD GPU Clock Tool. |
Send message Joined: 14 Dec 09 Posts: 161 Credit: 589,318,064 RAC: 0 |
I must state: Under room temperature my core clocks are 915/880 for 24/7. The above values (940/940) are done because i relocated the case next to an open window and the season is winter (Turkey). So, a room temperature of around 16-18 should provide the suitable environment. |
Send message Joined: 30 Dec 07 Posts: 311 Credit: 149,490,184 RAC: 0 |
Ah yes, winter helps the clocks. It was 35°C here today (Australia), my 5870 is a Vapor-X model so is coping OK even at 960 core clock. If I was still using my 4890 I would have needed to use w1.2 parameter today on MilkyWay or switch to Collatz. I like to contribute to both ATI projects, so I might swap back to Collatz for a while after tomorrow, now that my amount of pending has reduced. |
Send message Joined: 14 Dec 09 Posts: 161 Credit: 589,318,064 RAC: 0 |
To people shifting to server distributed ati application, it seem that the cpu load raised from 0.5 to 0.87 and i haven't noticed an increase in the performance. Can i increase it to 1 so that the each gpu core uses one cpu core? (In this case for 5970, two cores)? |
Send message Joined: 14 Dec 09 Posts: 161 Credit: 589,318,064 RAC: 0 |
I found a great spot: 970/1150 (core/memory) at 1.2/Default voltages. The values are for 24/7 since my window is open for 2 days:) and not even a single declock. (Thanks winter). I would go further but my 750 watt psu limits me. After a point, the memory gets much more important. I would say 1000 mhz would be fine also; but to guarantee any memory bottleneck, make it 1150. Hope people benefit by the info. ATI GPU 0: ATI Radeon HD5800 series (Cypress) (CAL version 1.4.515, 1024MB, 3104 GFLOPS peak) ATI GPU 1: ATI Radeon HD5800 series (Cypress) (CAL version 1.4.515, 1024MB, 3104 GFLOPS peak) NOTE: The total gflops value is tremendous. The longest wu duration is 1:14~15 |
Send message Joined: 17 Feb 08 Posts: 363 Credit: 258,227,990 RAC: 0 |
... I would say 1000 mhz would be fine also; but to guarantee any memory bottleneck, make it 1150. Hope people benefit by the info. Memory speed is not a bottleneck here. In fact it doesn't matter at all. You should try to reduce it to ~500 MHz, which should lower the GPU temp quite a bit, saves energy and might allow you to OC the core a bit further (which is the only important part here at MW... On Collatz it's the other way around) Join Support science! Joinc Team BOINC United now! |
Send message Joined: 14 Dec 09 Posts: 161 Credit: 589,318,064 RAC: 0 |
In fact it doesn't matter at all. For what gpu are you talking about? This is 5970, so please don't run arguments based on 5870. I have already been experimenting with the idea to lower the memory and to higher the core and if someone does what you just said, there will be no performance gain at all, just a waste of power and heat (at least for 5970). The minimum should be the around the same frequencies which is for example 900/900. Remember, this is a dual gpu. Not a single 5870 and what you're saying can't be true, at least the beast hardening under the hood does say your opposite. Also, the msi afterburner lets you down as far as 760 mhz... |
Send message Joined: 17 Feb 08 Posts: 363 Credit: 258,227,990 RAC: 0 |
In fact it doesn't matter at all. It applies to all ATI GPUs here.
You know that GP-GPU is != gaming ?
No need to remember that.
But it's a fact.
Huh ? Join Support science! Joinc Team BOINC United now! |
Send message Joined: 14 Dec 09 Posts: 161 Credit: 589,318,064 RAC: 0 |
It's your true and I don't intend to change your true. If i present another spot fore core/memory clocks, please don't state "better to lower the memory" or "the necessity of memory on Collatz" Thank you. "at least the beast hardening..." it was a silly metaphor for 5970...Never mind it... |
Send message Joined: 17 Feb 08 Posts: 363 Credit: 258,227,990 RAC: 0 |
It's your true and I don't intend to change your true. If i present another spot fore core/memory clocks, please don't state "better to lower the memory" or "the necessity of memory on Collatz" Thank you. OK... i give up on this one since i apparently have no clue what i'm talking about... LOL Join Support science! Joinc Team BOINC United now! |
©2024 Astroinformatics Group