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

[ao_ika_red]

I believe one day, somebody will modify its BIOS so it can utilise full Radeon Pro suport and probably will have 1/2 FP64 performance.

 

[Assimilator1]

Sounds good! , got any results yet?

Looking at the last 4 results from my 2nd rig (which stopped crunching a few days ago), it's seems their might be a new WU again! Unless it's just a fluke?
Anyone else seeing many 243.61 credit WUs?

1571998891728311840566187 24 Feb 2019, 18:25:49 UTC 27 Feb 2019, 0:20:24 UTC Completed and validated 71.32 10.22 227.16 MilkyWay@Home v1.46 (opencl_ati_101)
1562225311727825597566187 23 Feb 2019, 20:43:01 UTC 27 Feb 2019, 0:01:02 UTC Completed and validated 76.21 10.69 243.61 MilkyWay@Home v1.46 (opencl_ati_101)
1562094451727818411566187 23 Feb 2019, 20:25:06 UTC 24 Feb 2019, 22:28:23 UTC Completed and validated 78.43 10.97 243.61 MilkyWay@Home v1.46 (opencl_ati_101)
1558929301727666668566187 23 Feb 2019, 12:02:15 UTC 24 Feb 2019, 20:50:29 UTC Completed and validated 80.45 11.59 243.61 MilkyWay@Home v1.46 (opencl_ati_101)

Looking at crashtech's results from his no1 rig, whilst their are a lot of 243 credit WUs, their are many more 227.1x WUs, what I did notice (again) is the wide variation in run times, from high 80s to low 100s seconds.

Anyone else seeing this?

 

[IEC]

227.13 credit WUs appear to take ~21.05s on a Radeon VII. I don't see full GPU utilization on a TR 1920X. I might have to test the VII on my 8700K OC at some point to rule out a CPU bottleneck.

May also need to try running multiple WUs at once.

 

[IEC]

Running 4 WUs at once, dedicating 1 CPU core per task yields a task completion time of ~40-46s. Meaning this is almost twice as efficient (and shows in higher GPU temps).

I am actually able to run up to 6 WUs at once with a completion time around 60s, but this hits the hotspot max of 110°C and throttles the card a bit.

I think I need to water cool this puppy and run 6 WUs at once for maximum damage.

 

[Assimilator1]

Lol! Sweet! On the output that is!
Radeon VII?? That's a new one to me, looks like I need to do some reading.

What sort of spread of WU types (credit) are you seeing?

 

[Assimilator1]

Just looked at some valid tasks by uallas5, & currently it seems the most common credit WUs are 227.51 & .53, so they've changed again.
It seems that long term benchmarking is no longer possible as the WUs change frequently now.

 

[Assimilator1]

Cross posting from another thread with some MW@H benchmarks, will be interesting to see how long the 227.5x WUs stick around.....

@Assimilator1 , here are the 3700X data for Milkyway@home.

OS: linux mint 19.1
Computer: 3700X @3.6 GHz (8 cores/16 threads, SMT on), 2933 DDR4 14, 14, 14, 31.
APP: Separation v1.46 (227 credit)
Average run time (20 tasks): 4056 seconds
Power draw from the wall: 115 watts

It was a mix of 227.51, 227.52 and 227.53 tasks. I didn't see much variation in completion times.

So my i7 4930k @4.1 GHz running 12 MW@H threads is pulling ~232w from the wall (230v), average of it's 1st 12 tasks is 5378s.
That's a 50/50 mix of 227.51 & .53 credit results, varying from 5340-5474, all bar that latter one are from 5340-5387s, 227.51's mostly ranged from 5360-5370, 227.53 from 5367-5387 bar one.
So your rig's crunching WUs about 25% faster and using 1/2 as much power!!
Amazing! I know my rig's CPU is a few+ generations old, but even so, that's still impressive! I've got to get me a Zen 2 soon!

[edit] And I'd forgotten about the fact that your Zen2 has another 2 cores over my i7. So total output (for MW at least), would actually be about 58% higher! So even more impressive! (Obviously GPU crunching is faster for MW@H).

 

[StefanR5R]

Sadly, Radeon VII is reportedly no longer being produced; there is just remaining stock being cleared. Meanwhile, its server sibling Radeon Instinct MI 50 is apparently only available through server OEMs, and as for the fully enabled MI 60, AMD either never put up a product page for it on www.amd.com, or have taken such a page down in the meantime.

 

[Assimilator1]

Oh, what's it been replaced by? And how does their DP/FP64 compare?

 

[StefanR5R]

Radeon VII:

Oh, what's it been replaced by? And how does their DP/FP64 compare?

There is no direct replacement, and no known roadmap for one.

From the Zen2-any-interest thread:

But there is always the option to re-run select tasks outside of boinc.

Interesting, I've been thinking of needing something like that for benchmarking MW for a few years, and somewhere around 10yrs ago for SETI! How do you do that?

This needs to be figured out for each project, or rather each application, individually. So far I only researched this for PrimeGrid's LLR based subprojects:
Thread "PrimeGrid: CPU benchmarks", post #44

Since only my CPUs are busy for the time being, and the GPUs idle, maybe I can find a couple of spare hours to check out how to run MilkyWay tasks stand-alone.

 

[Assimilator1]

Looking at Crashtechs results for one of his R9 200/HD7900 rigs it seems the 227.5x credit WUs are still around (along with 227.12), maybe it could be worth collecting benchmarks again? What credit WUs are other people seeing?
(Btw Crashtech's rig above (835797) is doing 227.12 WUs in ~95-115s, 227.51/2/3 in ~85-100s, quite a variation with each WU! His RX Vega rig is doing 227.51 WUs mostly in ~150-160s).

 

[Assimilator1]

Ok, well I don't know how long I'll keep it up, but I'm going to start collecting benchmarks again, it seems the current WUs are more variable then previous ones though, even when the exact same credit WUs are looked at. So with that in mind I'll be posting & looking for a WU time typical range, excluding outliers, I've done it by eye from ~20 WUs (I ignored 2 that were at ~74s), but if you want to do properly with maths then be my guest .
I'll add it to the op if/when we get to several results or so. Post parallel WU times as well if you like, just LMK .

227.5x credit WUs, run time range for single WUs
RX 580 8GB (stock, 1350 MHz GPU, CPU Ryzen 3600, underclocked to ~3.7 GHz) ....................... 88-93s .... Me! [update] average seems ~90s

Surprised to see my card faster than a 7900! Would be interesting to know exactly what card it has? @crashtech Turns it out he runs 3 WUs at once.
R9 200/HD 7900 series card (comp 835797), (CPU AMD Ryzen 5 1600) ........................................ 90-100s .... crashtech (I looked through 2 pages of valid results, possibly running multiple WUs, TBD).

 

[StefanR5R]

Surprised to see my card faster than a 7900! Would be interesting to know exactly what card it has? @crashtech

@crashtech's account has been disabled for a while. He is in diaspora at TeAm AnandTech's secondary forum at http://teamanandtech.org/.

Is it beneficial to throughput to run more than one job at once on cards like @crashtech's? If yes, then he'll certainly making use of this.

--------

From the Zen2-any-interest thread:

But there is always the option to re-run select tasks outside of boinc.

Interesting, I've been thinking of needing something like that for benchmarking MW for a few years, and somewhere around 10yrs ago for SETI! How do you do that?

This needs to be figured out for each project, or rather each application, individually. So far I only researched this for PrimeGrid's LLR based subprojects:
Thread "PrimeGrid: CPU benchmarks", post #44

Since only my CPUs are busy for the time being, and the GPUs idle, maybe I can find a couple of spare hours to check out how to run MilkyWay tasks stand-alone.

I'll get to that eventually. But when? I don't know. It doesn't help that 6 GPUs are busy at Folding@home now, and 3 GPUs are down and waiting to get their waterblocks disassembled and cleaned. And the dayjob keeps me from doing much of this sort on workdays.

 

[Assimilator1]

Yea you get more ppd running at least 2 at once, (possibly more with high end cards IIRC), but it messes with WU times, so for benchmarking purposes it's better to run them singularly. And you're probably right he runs them in parallel, which would explain the WUs times being slower than my main rig's card (which I know from old MW benchmarks is slower than old my HD 7870 XT DS).

His account's been disabled? Is that what the line through his name means? Did he get into trouble?
I'll say hi to him there anyway :thumbsup:

 

[Assimilator1]

Well, I've noticed that quite a lot of the 227.5x WUs are also being done in 70-80s on my RX 580, so it seems MW is still no longer benchmarkable, aside from giving very broad times , oh well.

 

[Assimilator1]

An approximate time for the Radeon VII.

Sorry I missed this question. I'll have to give a fuzzy answer. I've been controlling high GPU temps by setting the fan speed at around 55% (tolerable noise for me) and setting power limits. I finally settled on a power cap of 150 watts for 24/7 use. I did run Milky Way some time ago but I don't remember what power cap I was using at the time. When I did test it, I was seeing 40 sec run times crunching 4 tasks simultaneously. I suspect that number will be higher at a 150 watt power cap.

 

[Assimilator1]

Tested with a MW wu, for @Assimilator1, got 16 seconds for one. Previously on Windows 3 at a time took 32 seconds, Mark(assimilator1).

Cheers Tony, that's rapid! What credit WU is it?

 

[TennesseeTony]

Credit is variable it seems.

 

[StefanR5R]

Managed to snag a XFX-branded Radeon VII at Newegg today.

I bought it mainly for the FP64 compute (3.46 TFLOPs i.e. 1/2 of MI50) and will update this thread when I get a chance to test it sometime next weekend.

I believe one day, somebody will modify its BIOS so it can utilise full Radeon Pro suport and probably will have 1/2 FP64 performance.

AMD themselves came out with a mod... ;-)
They revived Radeon VII as Radeon Pro VII with TDP reduced from 300W to 250W but FP64 fully enabled like on the MI50.
(AnandTech news item)
Availability mid June, MSRP $1,899.

 

[ao_ika_red]

AMD themselves came out with a mod... ;-)
They revived Radeon VII as Radeon Pro VII with TDP reduced from 300W to 250W but FP64 fully enabled like on the MI50.
(AnandTech news item)
Availability mid June, MSRP $1,899.

It's a true DP king, the only problem is price.

 

[IEC]

It's a true DP king, the only problem is price.

The Radeon VII does 1/2 the FP64 as the new Radeon Pro VII at ~1/3 the price. Potentially as low as 1/4 the price if you find the right used deal as it's only roughly equivalent to a RTX 2080 in gaming capabilities.

I'm currently using it in my main desktop since I have sold off nearly all my 12nm/14nm GPUs and just have a few left pending RMA or sale.

 

[Kiska]

The Radeon VII does 1/2 the FP64 as the new Radeon Pro VII at ~1/3 the price.

You sure the Radeon VII is 1/2? TPU seems to think otherwise... https://www.techpowerup.com/gpu-specs/radeon-vii.c3358
And Milkyway@home seems to think its 1/5:

Device 'gfx906' (Advanced Micro Devices, Inc.:0x1002) (CL_DEVICE_TYPE_GPU)
Board: AMD Radeon VII
Driver version:      3004.8 (PAL,HSAIL)
Version:             OpenCL 1.2 AMD-APP (3004.8)
Compute capability:  0.0
Max compute units:   60
Clock frequency:     1801 Mhz
Global mem size:     3221225472
Local mem size:      32768
Max const buf size:  3221225472
Double extension:    cl_khr_fp64
Build log:

Estimated AMD GPU GFLOP/s: 1081 SP GFLOP/s, 216 DP FLOP/s

 

[IEC]

@Kiska
1/2 of 1:2 is 1:4

Perhaps I should have written 50% instead of 1/2 as I can see how that might be confusing.

 

[Assimilator1]

Mmmm, would love to have that card for MW or Einstein

IEC
Did you ever get to water cooling your VII?

 

[Icecold]

I just set my Firepro S9150 back to run one task at a time instead of 4 for the purpose of this benchmark thread and ran some WU's:

Run time (sec)​	Credit​	Application​
33.16​	227.53​	Milkyway@home Separation v1.46 (opencl_ati_101) windows_x86_64​
31.13​	227.12​	Milkyway@home Separation v1.46 (opencl_ati_101) windows_x86_64​
31.14​	227.12​	Milkyway@home Separation v1.46 (opencl_ati_101) windows_x86_64​
30.17​	227.51​	Milkyway@home Separation v1.46 (opencl_ati_101) windows_x86_64​
31.26​	227.51​	Milkyway@home Separation v1.46 (opencl_ati_101) windows_x86_64​
31.14​	227.12​	Milkyway@home Separation v1.46 (opencl_ati_101) windows_x86_64​
31.19​	227.51​	Milkyway@home Separation v1.46 (opencl_ati_101) windows_x86_64​
34.16​	227.53​	Milkyway@home Separation v1.46 (opencl_ati_101) windows_x86_64​
27.14​	227.51​	Milkyway@home Separation v1.46 (opencl_ati_101) windows_x86_64​
31.14​	227.12​	Milkyway@home Separation v1.46 (opencl_ati_101) windows_x86_64​
25.32​	227.12​	Milkyway@home Separation v1.46 (opencl_ati_101) windows_x86_64​

The following are while running 4 WU's at a time:

Run time (sec)​	Credit​	Application​
96.54​	227.52​	Milkyway@home Separation v1.46 (opencl_ati_101) windows_x86_64​
83.47​	227.51​	Milkyway@home Separation v1.46 (opencl_ati_101) windows_x86_64​
87.9​	227.51​	Milkyway@home Separation v1.46 (opencl_ati_101) windows_x86_64​
95.56​	227.53​	Milkyway@home Separation v1.46 (opencl_ati_101) windows_x86_64​

If my math is correct - running 1 WU at a time comes out to around 641,000 PPD. 4 WU's at a time comes out to around 865,000 PPD.