Jump to content


Photo

Atomic Superpi Scores V5


  • Please log in to reply
787 replies to this topic

#781 Rybags

Rybags

    Immortal

  • Super Hero
  • 35,058 posts

Posted 29 May 2017 - 09:45 PM

Re yCruncher.

 

I've just been doing some side-by-side BMs with my Haswell i7-4790 (3.6-4.0) vs an Ivy Bridge i7-3770 (3.4-3.9).  Only a generation apart, in fact about 14 months.

With SuperPi they were almost the same, wPrime the earlier one was actually a touch quicker in the 4T 32M one though that was bare desktop vs open browser and other stuff going on.

 

The surprise packet is yCruncher though.  My Haswell almost twice as fast.  I suspected it might be writing to HDD which could severely handicap a mech vs SSD but it seems it doesn't.

Just goes to show the improvements in the maths processing as time goes on.  I've noticed it before with video encoding, the first gen i7s simply whip the arse of the Core 2 at similar frequencies.

Similar story with some types of work as you get newer i- generations.



#782 sora3

sora3

    Guru

  • Atomican
  • 13,900 posts
  • Location:At the bridge of the Kongo

Posted 03 June 2017 - 01:14 PM

Ryzen%20SuperPi%20Run_3.9_RAM_zps7wsfvlw

 

New AGESA has done massive wonders. Will try and push the CPU harder as I can probably nail 4GHz at this rate...


Main Rig: AMD Ryzen 7 1700 @ 3.9GHz/Asus X370 Prime/Corsair DDR4 3000 32GB @ 2667(!)/Sapphire RX Vega 64/WD Green M2 240GB/Bitfenix Shinobi/Noctua NH-D14/Corsair HX-750

 

Camera rig: Sony Alpha a77/Sony DT 16-50mm f2.8 SSM/Sony DT 16-105mm f3.5-5.6/Tammy 70-200mm f2.8/Sony DT 50mm f1.8/Sony HVL F42AM Flash


#783 Rybags

Rybags

    Immortal

  • Super Hero
  • 35,058 posts

Posted 03 June 2017 - 01:35 PM

How does it go on yCruncher?  Try tests (0,1,2) through to (0,1,5).



#784 sora3

sora3

    Guru

  • Atomican
  • 13,900 posts
  • Location:At the bridge of the Kongo

Posted 03 June 2017 - 07:31 PM

Well, will try now. :D

 

EDIT: y-cruncher_Ryzen_3.9_zpspihm3gjw.jpg

 

Because Rybags wanted it...

 

y-cruncher_Ryzen_3.9_simple_zps9oxlcske.

y-cruncher_Ryzen_3.9_harder_zpsxdvg8cfe.

y-cruncher_Ryzen_3.9_blah_zpswgmlsvw2.jp

y-cruncher_Ryzen_3.9_blah2_zps0k71cqiu.j


Edited by sora3, 03 June 2017 - 07:40 PM.

Main Rig: AMD Ryzen 7 1700 @ 3.9GHz/Asus X370 Prime/Corsair DDR4 3000 32GB @ 2667(!)/Sapphire RX Vega 64/WD Green M2 240GB/Bitfenix Shinobi/Noctua NH-D14/Corsair HX-750

 

Camera rig: Sony Alpha a77/Sony DT 16-50mm f2.8 SSM/Sony DT 16-105mm f3.5-5.6/Tammy 70-200mm f2.8/Sony DT 50mm f1.8/Sony HVL F42AM Flash


#785 mark84

mark84

    Champion

  • Hero
  • 5,723 posts

Posted 29 June 2017 - 02:11 PM

Nice Sora. Updated your SuperPi and ycruncher scores in OP

 

Re yCruncher.

 

I've just been doing some side-by-side BMs with my Haswell i7-4790 (3.6-4.0) vs an Ivy Bridge i7-3770 (3.4-3.9).  Only a generation apart, in fact about 14 months.

With SuperPi they were almost the same, wPrime the earlier one was actually a touch quicker in the 4T 32M one though that was bare desktop vs open browser and other stuff going on.

 

The surprise packet is yCruncher though.  My Haswell almost twice as fast.  I suspected it might be writing to HDD which could severely handicap a mech vs SSD but it seems it doesn't.

Just goes to show the improvements in the maths processing as time goes on.  I've noticed it before with video encoding, the first gen i7s simply whip the arse of the Core 2 at similar frequencies.

Similar story with some types of work as you get newer i- generations.

Yeah the guy that makes ycruncher makes bespoke binaries for each architecture to make the best use of the intruction sets and hardware accelerated logic available.

Even on the same architechure but with a newer version of ycruncher I've seen large performance gains as he improves his algorithms and tries to make the calculations happen as efficiently as possible with the hardware resources available.

 

Which is a better way of benching imo, it delivers the real full potential of the chip rather than running the same thing on every CPU the same way, which might not be the best or fastest way of executing it. Like SuperPi and it's long outdated x87 intructions, newer and better methods come along over time. Doing a 1M superpi quivalent run in ycruncher (0, 1, 0, 20), my work (Zen) PC can do the math in 0.1s as opposed to 11.9s.

 

RE HDD writing, the ycruncher compute time is before it writes the results to disk, so that doesn't affect the score. Only the start-to-end wall time is affected by the mass storage sub system.

 

http://www.numberwor...on_history.html

Always a good read.

If you see his March 2017 update, most binaries copped a +5% speedup, Zen was added and got a 14% increase too.


Edited by mark84, 29 June 2017 - 02:17 PM.

Desktop: i7 6850K @ 4.3GHz | Titan X (Pascal) | 32GB RAM | 120GB + 500GB SSDs | 2x 5TB HDD in RAID1

Server: G3258 @4.6GHz | 120GB SSD | 2x 4TB RAID1 | 2x 6TB RAID1

Desktop build log | Firefox AtomicMPC Menu | The Truth About Graphics Power Requirements V2 | CPU bench league

www.atomicsimracing.net - The place for Atomicans who love motor sports and sim racing.


#786 Rybags

Rybags

    Immortal

  • Super Hero
  • 35,058 posts

Posted 29 June 2017 - 02:19 PM

I don't see the value in platform specific enhancements, really the bench should be generic and just use features widely available. 

 

I also don't like versioning where results change, all that serves to do is invalidate months worth of data.  The value of legacy albeit a bit obsolete benches like SuperPi, wPrime is that they went through only a few versions once becoming popular and the results expected didn't deviate too much.

 

The value of these benches shouldn't be eWank value but the ability to make relative comparisons which then can be basis for good decisions relating to upgrading or buying new gear.  If results are skewed or handicapped or vary depending on which version was run, then that particular benchmark loses it's value.



#787 mark84

mark84

    Champion

  • Hero
  • 5,723 posts

Posted 30 June 2017 - 05:21 PM

Well yes. But if you're going to be calculating Pi (or anything) in 2017 for sciencetic stuff are you going to be using x87 or the latest accelerated instruction sets?

 

You make a good point regarding comparitive results. And of course that's needed. But equally you want benchmarks to reflect real world capability, if you can exploit AVX-512 on one CPU to make it legitimately process something faster you're not showing it's full real world potential by not using it. They may be comparitive, but not the whole story. There's room for both methods imo


Desktop: i7 6850K @ 4.3GHz | Titan X (Pascal) | 32GB RAM | 120GB + 500GB SSDs | 2x 5TB HDD in RAID1

Server: G3258 @4.6GHz | 120GB SSD | 2x 4TB RAID1 | 2x 6TB RAID1

Desktop build log | Firefox AtomicMPC Menu | The Truth About Graphics Power Requirements V2 | CPU bench league

www.atomicsimracing.net - The place for Atomicans who love motor sports and sim racing.


#788 Rybags

Rybags

    Immortal

  • Super Hero
  • 35,058 posts

Posted 30 June 2017 - 05:43 PM

The point of getting a value of Pi itself is a bit irrelevant.  It's a benchmark for benchmark's sake, not for mathematical sake.

To be worthwhile you have to just pick a landmark set of CPU features and stick with that.  If you favour a new set of features that only certain CPU models support then the results can be unfairly skewed.

 

But in the real world you need benches that can reflect real world results.  Generally number-crunching benches don't do that entirely well.  They're affected by how well CPUs do FP work, and I've also found since they're so intensive that they're also affected greatly by decent memory bandwidth.  You can get a greatly variable result e.g. in yCruncher by a pair of Core i CPUs that are a couple of generations apart but otherwise very close for other tasks.






0 user(s) are reading this topic

0 members, 0 guests, 0 anonymous users