MilkyWay@H - Benchmark thread Winter 2016 on (updated 1-2021) - GPU & CPU times wanted for new WUs

[Assimilator1]

Bloody hell! I've just noticed your Thunder Strike rig, 56 threads!! Awesome!

Anyway, re your R9 280Xs, so if I understand correctly you have 2 of them, they were both in the Xeon rig running @ 1GHz, & then you moved each of them to different Pentium rigs & different GPU clocks?
I assume the Xeon has much larger cache? In which case it's also interesting to see that clock speed counts a lot more than cache size.

 

[TennesseeTony]

No, I have three systems containing 280X's, 5 active cards, and I just found a sixth (and seventh) card under a pile of rubbi......errr, under some carefully archived elder computer parts, and one of them works! This sixth card runs at 1080GHz, which is why I began the search, as my times on the first page/post list at 1080GHz, so I knew something was awry, as my currently installed cards are either 1000 or 1100MHz.

Of those 3 systems, two of them are seldom run (contests mostly: Milkyway, Einstein, Enigma and hopefully Universe GPU some day).

https://forums.anandtech.com/thread...elp-naming-a-rig.2441945/page-4#post-38874541

So, only the thread count caught your eye, regarding Thunder-Strike?

 

[Assimilator1]

Wow! That's PC porn! Lol

So are you running that 280X @1080 GHz on liquid Helium? Any problems sourcing the Helium? How near to absolute zero can you get?
Nice find on the spare 280X

Back to MW, looking at a few more 227.26 credit WUs I've found that they take about 56s vs the 227.23s of ~53s, so the .26s take about 5% longer, on my HD 7970 at least.

 

[Assimilator1]

Added 2 more GPU times, 2 more to come when I get full details

GTX 1060 3GB 250s - DVDL - CPU TBC

 

[Assimilator1]

Table updated with additional times, including a Vega 56! , seems pretty good at MW for a modern card!

 

[StefanR5R]

(from the Folding@Home Holiday Season Race thread)

Oh, FAHBench is back in AnandTech's GPU reviews.

Titan V's SP performance should be much better than 1080Ti. Driver issue? One thing for sure, this card will be the new king in Einstein, Enigma, and Milkyway.

It looks like Milkyway's opencl_nvidia application doesn't really care for FP64 on GPUs. Otherwise, the original Titan's and the Titan Black's relative score should be a lot higher.

Edit:
On the other hand, this list does not take into account the throughput that you can get when several tasks run in parallel on one card.

 

[crashtech]

..

(from the Folding@Home Holiday Season Race thread)

It looks like Milkyway's opencl_nvidia application doesn't really care for FP64 on GPUs. Otherwise, the original Titan's and the Titan Black's relative score should be a lot higher.

Edit:
On the other hand, this list does not take into account the throughput that you can get when several tasks run in parallel on one card.

I didn't think there could be a fundamental difference in FP64 demand if the application requires that level of precision. I have read that drivers can expose functionality that reduces the number of EUs to perform DP tasks, if the hardware exits on the card to do so (Titan Black, for instance).

 

[Assimilator1]

European Unions??

Stefan
Yea MW's NVidia apps don't seem to use their DP performance to anywhere near their optimal (unless multiple tasks are run), GPU load on an NVidia running a single task is nowhere near 100%.

 

[ao_ika_red]

Is it like AMD cards unable to utilise their SP capability even though theoretically they had same SP performance with Nvidia counterparts?
Oh, is this benchmark thread still relevant? If so, I'll post my score.

 

[Assimilator1]

Re AMD SPs, no idea, I'm behind the times on modern cards atm.

Re benchmark, as per the op, yep!

 

[ao_ika_red]

Well then. This is mine.
AMD RX 560 2GB @1300MHz on Athlon x4 845.
It took 238.83s to finish a 227.23 credited task.

Spoiler: Milkyway Run

 

[crashtech]

Looking at the theoretical DP performance of a Titan Black, it seems they should perform about 1.7x as well as a 7970 GHz Edition. Any word on how they do? Prices have been falling on them, and they seem a likely candidate for a dual purpose card, since their stock SP perf is roughly in line with a boosted 980.

 

[StefanR5R]

http://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=4171

The key is to run seven (7) task concurrently since approximately 14% of the TITAN GPU core is used for one MW task. Running seven (7) at the same time with DP enabled gets you around 99% GPU utilization and then will the card start to match or outperform a 7970 or 280x card.

I think he means either the original Titan, or Titan Black.

 

[crashtech]

In theory it should do better, but just as AMD's theoretical SP performance does not always align with actual performance, Nvidia's DP seems to have a similar problem. I see Titan Blacks going for $275, which is probably not terrible, since it's stock SP exceeds a boosted 980 by a fair amount. I wonder how they OC?

 

[Assimilator1]

Well then. This is mine.
AMD RX 560 2GB @1300MHz on Athlon x4 845.
It took 238.83s to finish a 227.23 credited task.

As per the op, if you read it , I need an average of 5 WUs, so as to reduce variation.

 

[panamaatx]

Multiple WU effective single process times?
Where can I find the effective fastest WU time per GPU running multiple WU? Would like to know the best WU yield with multiple parallel tasks.

 

[iwajabitw]

Multiple WU effective single process times?
Where can I find the effective fastest WU time per GPU running multiple WU? Would like to know the best WU yield with multiple parallel tasks.

The new MW work units are actually 5 in 1 so there really isn't a need to run multiples. On my 280x or 7970 1 task takes about 64 seconds. I've ran 2 at a time and have seen a slight gain in ppd just because the card isn't idle when finishing a unit. Your card may yield a different result. Check out the Number Crunching section at the MW message boards.

 

[panamaatx]

The new MW work units are actually 5 in 1 so there really isn't a need to run multiples. On my 280x or 7970 1 task takes about 64 seconds. I've ran 2 at a time and have seen a slight gain in ppd just because the card isn't idle when finishing a unit. Your card may yield a different result. Check out the Number Crunching section at the MW message boards.

Thanks for sharing. I've actually gone to the MW forum, but I'm not able to ask any questions since I'm only able to crunch on the GRCPool.

About how many daily GRC can I expect to generate at maximum output from one card with your specs above?
From some older posts I was given the impression one 7970 should yield 50GRC daily, but that hasn't been my experience.

Thanks for the info.

Panama

 

[iwajabitw]

Thanks for sharing. I've actually gone to the MW forum, but I'm not able to ask any questions since I'm only able to crunch on the GRCPool.

About how many daily GRC can I expect to generate at maximum output from one card with your specs above?
From some older posts I was given the impression one 7970 should yield 50GRC daily, but that hasn't been my experience.

Thanks for the info.

Panama

Sorry I have no experience with GRC.

 

[crashtech]

The new MW work units are actually 5 in 1 so there really isn't a need to run multiples. On my 280x or 7970 1 task takes about 64 seconds. I've ran 2 at a time and have seen a slight gain in ppd just because the card isn't idle when finishing a unit. Your card may yield a different result. Check out the Number Crunching section at the MW message boards.

Wow, this is news to me. So the new sweet spot for a 7970/280x is two tasks now? I settled on three last year and haven't revisited it, thanks for the heads up!

 

[iwajabitw]

Wow, this is news to me. So the new sweet spot for a 7970/280x is two tasks now? I settled on three last year and haven't revisited it, thanks for the heads up!

For marathons running 1 task seems to be good enough, for Sprints I would switch to 2. In 2017 MW started changing the work units to help take the load off of there server. Thats why the points went up to 227.xx per task. Something like that anyway.

 

[crashtech]

For marathons running 1 task seems to be good enough, for Sprints I would switch to 2. In 2017 MW started changing the work units to help take the load off of there server. Thats why the points went up to 227.xx per task. Something like that anyway.

A quick check on my 280x showed it averaging 2:03 per task, 3 at a time. This to me means it is ready to upload one about every 41 seconds. When I changed it to 2 concurrent tasks, average completion times were 1:32, which is ready to upload one every 46 seconds. It wasn't a comprehensive test, but this points to 3 tasks concurrent still being a bit faster, unless I am figuring it wrong.

 

[iwajabitw]

It may Crash, its been a while since I tried, at least before the FAH race. Wife said the house felt cool last night so I fired up Einstein on a few rigs.

 

[StefanR5R]

Earlier in 2017 when I ran the 133.66 credits WUs, the 5 parts were done serially (one after another). Did they change that?

I suspect they didn't. Just checking with CPU tasks and GPU tasks on Linux: The progress indicator of each task includes steps that are multiples of 20 %. Running one task at a time on 1080Ti, core utilization fluctuates between 75...85 %.

 

[Assimilator1]

Still done serially IIRC, just the progress bar doesn't reset for each task.

NVidia cards in particular benefit from running multiple WUs at once.

paanamaatx
What's GRC??