My 1090T is beating my 2700x on ECM... any ideas why?

Alles zum Projekt yoyo@home
Everything about the project yoyo@home
Nachricht
Autor
mister.marmot
PDA-Benutzer
PDA-Benutzer
Beiträge: 34
Registriert: 15.07.2016 00:53

My 1090T is beating my 2700x on ECM... any ideas why?

#1 Ungelesener Beitrag von mister.marmot » 15.02.2021 02:36

The 1090t is overclocked to 3625mhz (135Watts) w/ RAM underclocked to 1600@1.35v (couldn't get the timings right at 2000/2200mhz and gave up till later). The RAM throughput is 9.5Gb/s (single core) and 13.2Gb/s (multi)
The 2700x is overclocked to 3800mhz (99 Watts) w/ RAM overclocked to 3200@1.50v. The RAM throughput is 29Gb/s (single) and 44Gb/s (multi).

They both have 16GB RAM and the 1090t is running 6x ECM (full cores) while the 2700x is juggling 8x COVID19 WCG, 4x Kryptos@Home VM's and 4x ECM (YoYo).

How is the 1090t outperforming the 2700x?

I set the machines to be visible to everyone and here are the links:

1090t https://www.rechenkraft.net/yoyo/show_h ... tid=340032

WU's
largest 845 credit: 22750 sec
medium 684 credit: 17750 sec
smaller 365 credit: 10000 sec
smallest: 269 credit : 11000 sec

2700x https://www.rechenkraft.net/yoyo/show_h ... tid=496714

WU's
largest 845 credit: unknown sec
medium 684 credit: 18800 sec
smaller 360 credit: 10000 sec
smallest: 269 credit : unknown sec

My hypothesis:
A) The WU's are large enough (on the 2700x) to force use of the swap file? (However, Process Hacker shows the machine has not gone above 14GB commits: real RAM = 16GB)
B) The WU's are in competition for the LVL 3 cache with the WCG and Kryptos WU's and being forced into DDR4 RAM usage during calculations?

I'm guessing option B) is most likely but you all understand the algorithms better than I; so asking you.

It would explain the 684 credit 17750s vs 18800s BUT I would think the 365 credit (smaller RAM footprint?) WU's would run faster on the 2700x but are coming in at an average of 10000 sec on both machines.

Benutzeravatar
n3Ro
FAQ-Leser
FAQ-Leser
Beiträge: 166
Registriert: 01.01.2019 13:01

Re: My 1090T is beating my 2700x on ECM... any ideas why?

#2 Ungelesener Beitrag von n3Ro » 17.02.2021 08:20

It is hard to explain your conclusion because it might be completely wrong. Please try to compare runtimes for the case where your 2700X runs 8 ECMs only and nothing else.
Hyperthreading is no replacement for real cores. If you run tasks in parallel on real cores, each core has (really simplified) dedicated execution units. If you run tasks on hyperthreading virtual cores, these share resources and therefore execution time with the actual real cores. The advantage of HT only comes into play when one task on a real core stalls (e.g. waiting for memory to be read), then the other HT Task can use the same execution resources and you see some performance gain. Otherwise without HT these would idle. This will never be the same as an additional core.
:zzz:
Two and a Half Men - Nur weil ich sage, dass ich verstehe, heißt das nicht, dass ich zugehört habe.

mister.marmot
PDA-Benutzer
PDA-Benutzer
Beiträge: 34
Registriert: 15.07.2016 00:53

Re: My 1090T is beating my 2700x on ECM... any ideas why?

#3 Ungelesener Beitrag von mister.marmot » 18.02.2021 07:02

n3Ro hat geschrieben:
17.02.2021 08:20

Hyperthreading
That's what I forgot about my new 2700x; it's 8 real/16ht.

The HT times are about 9% longer than the 1090t's real core times so the 2700x is doing 83% more work per core.

I'm not going to turn off HT during the winter months. Maybe when it's get's hotter I'll run a real core to real core BM.

My guess is the 2700x would do one of the 540 credit WU's in 13450 sec HT off.

Thanks for the reminder; all you had to say is "HT is on".

BTW, this got me to attempt more OCing and the 2700x RAM is now up to 3400Mhz and CPU is stable at 78C running 4000Mhz at 119W/1.33 VCore max

Antworten

Zurück zu „Number crunching“