Ryzen 7000 Series Efficiency

Andrew_Carr

2[H]4U
Joined
Feb 26, 2005
Messages
2,669
I picked up a 7950x to do some testing with. With the node shrink I was hoping for something like 20% gains in efficiency vs ryzen 5000/3000 series CPUs. So far these are my results:

Bottom line up front (Should I buy this for mining?): No

System info:

Motherboard: Asrock x670e Lightning (~$270)
Memory: G.skill flare x5 DDR5-6000 running at CL37 (advertised timings are CL36-36-36-96 @ 1.35V but motherboard XMP sets them to 37). PN F5-6000J3636F16GX2-FX5 ($220)
CPU: 7950X ($699+tax)
HSF: Noctua NH-D15 w/ kryonaut
PSU: Seasonic Focus 650W (Gold) All power readings are from the wall on a 120V line. Slightly better efficiency if using a titanium PSU on 220V

Test environment: latest version of HiveOS (0.6-219@220925) with the latest version of xmrig (v.6.18.0). Linux kernel is 5.10.0. Ambient temps around 74F in an open case. I'm not getting temp reporting data yet but let's assume 95C for everything based on recent 7000 series info. Computer only has a USB thumb drive as a hard drive, nothing else connected (no keyboard, mouse, monitor, SSD, wifi, RGB, etc.). 3 x 120MM case fans (low RPM). I think this represents a mostly ideal environment.

Testing methodology: Reduce PPT and record wall-meter values and hashrates (local, NOT pool) after a few minutes of stability.

Monero
Efficiency kept improving down to about 140W before tapering off. I need to come up with a better model for finding the ideal point to choose for power levels, since profitability of the coin & power cost heavily impact how efficient you want to be (ie. if profitability is high and electricity costs are low, then running at 303W might actually make more sense than 127W in extreme scenarios). Overall 130-150W (90W PPT limit) seems to be the ideal range currently
Raptoruem
I didn't have time to do multiple runs for 24 hour periods, but since xmrig doesn't require much tuning time and generally speaking if you know the peaks, lows, and some of the middle of the road hashrates you can estimate a good idea for where it will end up I think these are good approximate values. I tried to record a few different runs to get a rough idea. The only thing I've seen drastically improve hashrates on non-RandomX algorithms is L3 cache, which sometimes hugely boosts efficiency (~50% gains). So if 7000 series come out with 3d v-cache models for not a huge mark-up vs their same-core cousins a 16-core version with more l3 cache might be interesting. (The 5800X3D is kind of a flop for mining because it's expensive and it's only 8-core, but efficiency-wise it's up to 50% better in some algorithms than the equivalent 8-core 5000/3000 series CPUs; so if the price were right it would be a great buy.)
1664334387280.png


Final thoughts: I think I still have room to test with for RAM tweaks and such, but initial thoughts are it's totally useless for mining and the efficiency gains aren't panning out. My testing with Epycs and such makes me think that maybe by drastically lowering DDR5 throughput to benefit memory timings could help. At a certain point extra bandwidth doesn't matter (on Epycs more than 2 lanes of DDR4 at low speeds didn't seem to matter). Maybe people will find a better way to force the processor into an eco mode instead of using PPT limits (I have no eco mode on my motherboard apparently), but overall it seems like it's not much different than prior generations despite the huge cost increase. This was the same story with the 5000 series, where they only offered minor efficiency gains (except for the 5800X3D on certain algorithms) vs the 3000 series but at much higher cost. I generally don't like tweaking memory and going for expensive memory on CPU mining builds because the added cost (time & RAM cost) usually extends ROI way too much, but if there's something simple like set FCLK to X and making very rough timing changes at the expense of speed work then maybe I'll consider it. I'm not hoping for a silver bullet though and plan to stick to 3000 series for now.



Raw data recorded:
Everything stock except XMP on for the RAM (raw data):
Monero: 21.5khash/s @ 303W
Raptoreum: 4.7-5.3khash/s @ 160-180W (1.6khash min, 24.2khash max)

PPT @ 110W:
Monero: 19.5khash/s @ 163W
Raptoreum: 4.3-4.7khash/s @ 159W (didn't run long enough to see max)

PPT @ 90W:
Monero: 18khash/s @ 139W
Raptoreum: N/A (didn't record)

PPT @ 80W:
Monero: 16.9khash/s @ 126.6W
Raptoreum: 3.4 - 3.6khash/s @ 125.7W
 

Attachments

  • 1664334331102.png
    1664334331102.png
    38.8 KB · Views: 0
Last edited:

pututu

[H]ard DC'er of the Year 2021
Joined
Dec 27, 2015
Messages
2,529
Bottom line up front (Should I buy this for mining?): No
Since you have 7950x. Do some test for DC folks here ;)

Maybe try primegrid first. Subprojects such as Sophie Germain Prime Search SGS or Proth Prime Search Extended LLR PPSE. Each task should run very quickly, <20 mins. IIRC, tasks should have similar run time for the same search range. This will use AVX instruction code, so best to run 50% cpu (16 threads). Wanna see if AMD avx-512 has any big advantage, if the project can take advantage of this new instruction set.

P.S. might want to post DC stuff in DC forum and mining stuff here.
 

Andrew_Carr

2[H]4U
Joined
Feb 26, 2005
Messages
2,669
Since you have 7950x. Do some test for DC folks here ;)

Maybe try primegrid first. Subprojects such as Sophie Germain Prime Search SGS or Proth Prime Search Extended LLR PPSE. Each task should run very quickly, <20 mins. IIRC, tasks should have similar run time for the same search range. This will use AVX instruction code, so best to run 50% cpu (16 threads). Wanna see if AMD avx-512 has any big advantage, if the project can take advantage of this new instruction set.

P.S. might want to post DC stuff in DC forum and mining stuff here.
Kind of a busy week but I'll try to run some stuff this weekend. As long as disk IO doesn't matter it shouldn't be an issue. I'm going to try voltage offsets but I doubt it'll change anything.
 

Andrew_Carr

2[H]4U
Joined
Feb 26, 2005
Messages
2,669
Bench Verus for me please when you have a moment.
Sorry, missed this. Not having any luck with it at the moment but I'll play with it some more and see.

[Looks like roughly 50-64MH/s with the most efficient settings only getting around 50MH/s on verus - RabidMining tested this].
 
Last edited:
Top