Efficiency with Ryzen 5900x, 128GB ram and 2TB ‎NVMe (FireCuda 520)

[adamsky-01]

Did somebody test this plotter with Ryzen 5900x? What could be the expected time per plot with the below configuration:

• Ryzen 5900x
• 128 GB 4 x 32 GB DDR4 3600 MHz
• 2TB ‎NVMe Gen4 (FireCuda 520)

I'm just considering the ram upgrade to128GB.

UPDATE:
RAM Disk works pretty well for me. I'm running windows v0.0.6 version of plotter (256 bucket, 24 threads) with 112GB assigned to RAM Disk.
Total plot creation time: 28 min

[casa10]

Same question with 64GB of RAM (3600MHz)

[delucca]

I have a Ryzen 5900x with 32GB of RAM and a 970EVO

I'm taking 1h to plot k32

Since this plotter optimizes your plotting speed, it does not make sense to run parallel plots. So, I don't see any benefits in large RAMs, unless you use RAM as your tmp2 dir. The same applies to large NVMe

[adamsky-01]

tmp2 needs exactly 104 GiB, 128 GiB RAM is enough for about 12 threads, so it could just work out by a hair.

Thank you for the clarification. Most of the consumer grade motherboard is limited to 128 GiB of ram. This means only 12 threads can be utilized, so saving NVMe with 104 GiB ram disk and max 12 CPU threads or probably faster plotting with all threads but no ram disk benefit.
Will this limit of 12 threads (with 104 GiB ram disk and 128 GiB memory) significantly increase the plotting time? What time could be expected in early mentioned configuration with Ryzen 5900x (12 threads used only)?

[casa10]

i don't understand all.. Sorry..not clear from beginning for me.

• you can use your ddr RAM like temp2 directory. In that case, 104GiB is necessary. If you use 12 threads, you need less than 4GiB of RAM so 128GiB it's ok because you consider for example OS take 4-6Gib of RAM.
• if you have only 64 GiB of RAM (or less 32GiB etc.), can you configure tmp2 directory in this application ? Does she will write step by step from ddr RAM to SSD (or HDD) for the 104 GiB and used in same time the 220 GiB of space in the SSD ? In this case of use, do you write less in the SSD in total than using swar manager or plot manager where 1.8 TB of the SSD is using ?

[marctanw90]

weird. i threw 24 cores on my 5900x but its CPU usage is fluctuates a lot and averages around 33%

[casa10]

i tried with 128GiB of RAM tmp2 (ddr ram) and used tmp1 SSD.
40 min to make 1 plot with 12 threads.
CPU : 72°C max
EDC : 100% for all phase P1, 86-88% phase P2, 79-93% phase P3
Consumption : 102-115W for P1,90W for P2 , max 105W for P3
PPT : max 100% but often between 90-95%
I don't know if for CPU it's good to work like that often and 24/24hours.

[adamsky-01]

CPU : 72°C max

You are absolutely fine with this temp. on your CPU even 24/7 over years.

[thiagottjv]

Hi, I have a 5900x, 32Gb of RAM and 2 970 evo 2Tb (one for each temp destination).
Im doing one plot with around 1750s (30min). Each time a plot ends this time increase. With that said, Im using only around 16 to 17Gb of ram using all 24 threads. I`ll try to decrease the buckets to increase the ram usage and hopefully decrease the time.

[adamsky-01]

I run a quick test under Windows 10, no other tasks just single plot using madMAx43v3r/chia-plotter windows build v0.0.4
chia_plot.exe -n 1 -r 24 (5900X) -t (Samsung 970 EVO 500GB NVMe) -2 (FireCuda 520 2TB ‎NVMe Gen4 )

Phase 1 took 1168.58 sec
Phase 2 took 525.656 sec
Phase 3 took 694.047 sec
Phase 4 took 142.096 sec
Total plot creation time was 2530.42 sec (42 min)

It could be probably faster under the Linux. I'll update these results with the RAM disk and under Ubuntu when I'll receive the 128GiB RAM kit.

[adamsky-01]

Changing buckets from default 128 to 256 helped me to save 5 minutes on plot. Setting as above, still on Windows 10.

Phase 1 took 881.111 sec
Phase 2 took 510.704 sec
Phase 3 took 726.681 sec
Phase 4 took 104.45 sec
Total plot creation time was 2223.04 sec (37 min)

[thiagottjv]

Yes, seams that the way to go is increase the bucket size... With 128 first plot time was at around 1750s, with 64 it wont finish because I didn't had enough ram, but the phase 1 timings were higher.
With 256 I got 1630s, with 512, 1580s! So I went down almost 3min!
I`ll try to get higher.
The next test will be doing a raid 1 with my two nvmes and using the same folder for temp1 and 2.

[adamsky-01]

I've finished test with 64 buckets but the result was very similar to 128 buckets. Below results for the direct comparison.
OS: Windows 10
Plotter: chia-plotter windows build v0.0.4
CPU: Ryzen 5900X
RAM: 48GiB 3200MHz CL16
tmpdir: Samsung 970 EVO 500GB NVMe
tmpdir2: FireCuda 520 2TB ‎NVMe Gen4

64 buckets
Phase 1 took 1182.35 sec
Phase 2 took 520.179 sec
Phase 3 took 702.441 sec
Phase 4 took 96.3558 sec
Total plot creation time was 2501.42 sec (42 min)

128 buckets
Phase 1 took 1156.46 sec
Phase 2 took 513.086 sec
Phase 3 took 736.826 sec
Phase 4 took 103.345 sec
Total plot creation time was 2509.79 sec (42 min)

256 buckets
Phase 1 took 881.111 sec
Phase 2 took 510.704 sec
Phase 3 took 726.681 sec
Phase 4 took 104.45 sec
Total plot creation time was 2223.04 sec (37 min)

[adamsky-01]

Update:
RAM Disk works pretty well for me. I'm running windows v0.0.6 version of plotter (256 bucket, 24 threads) with 112GB assigned to RAM Disk.
Total plot creation time: 28 min

[jacek900]

For your configuration - with ram disk - the total time should be around 29 minutes for plot total. Mine is 1788 seconds but my memory is slower.
I am going to test RAID0 for temp1 to see if it speeds up phase 2 as my cpu seems to be a bit bored during that one.

[jacek900]

This is correct.

[adamsky-01]

Jacek madMAx43v3r mentioned in the post below that once you use 104 GiB of 128 GiB RAM for minimum size of RAMdisk you might be able to utilize 12 maybe 16 threads only due to low remained memory. Did you notice any issues running 24 threads or did you use any trick to to get around it?
#150 (reply in thread)

[marctanw90]

ram disk 104 is pushing it. I reckon go 110 and leave the rest for system / buckets

[jacek900]

Running on all 24 threads and haven't noticed any issues so far. 128GB total system ram and 110GB ram disk.

[vctrb]

My 3900X does lab runs in 1600 seconds, production runs in 1660 seconds.

Your nvme is not made for this job.

[casa10]

Some returns about CPU temperatures according each phases?

[adamsky-01]

On the left current temps, on the right max. recorded temps.
I didn't check temps for each phase.

[adamsky-01]

It's almost silent setup. Plenty large, low spinning fans, Noctua NH-D15 coolers. Temps. are acceptable, the thermal throttling starts in somewhere between 90c and 100c so still way below 90deg.

[casa10]

Thank you.
Temperature often more than 70°C. After normal because if you use all threads or even half, processor will burn to make all the work.
Fortunately lik you said less than max autorized.

[leonidbkh]

I made last plot in 1700s on my 5800x with 110gb ramdisk and 1tb 970 evo plus as first temp. Still have some problems with CPU cooling, need to change cooling system, have 90 degrees at 1st phase