it is gen 3 now but nothing really changed, I tried the Nvidia A2000 too but the same as well.
I guess my plans of using the A2000 as an efficient alternative for farming (not plotting) seem to be ruined.
Also, I doubt this has anything to do with hardware issues, at this point, it looks like a software-related issue.
Edit: looking at the HP Z420 docs, both x16 slots are Gen3 and should work just fine (theoretically it should do ~ 15GBs)…I will try to get my hand on a 3060Ti to make it an exact replica of @madMAx43v3r 's testing/dev machine.
Edit2: why is phase 2 so fast compared to other phases? and it also seems to be using the whole PCIe bandwidth.
Not sure if this information is interesting to you or not but figured I’d share.
I have a Z420 with an E5-2670 and 256GB of 1866Mhz memory.
-GTX 1060 6GB with -S 3 so I don’t run out of Vram I get 10 min plots
-RTX 3060 I get about 4 minute plots
I also have a Ryzen 5900X system with 128GB of Memory
-RTX 3070TI plotting with dual Samsung 980 Pros in raid 0 I’m right around 2 minutes
Nice! this is the kind of info I was looking for! so basically the results I am seeing right now are typical, this means I don’t have to waste my time anymore debugging.
Edit: however the Nvidia A2000 results seem too low to me, the A2000 is supposed to be way better than what I am getting right now.
Plotter is RAM bandwidth sensitive, so you need to be careful when using 2 CPUs as NUMA may wreck havoc there. I would start with just one CPU (remove the other), as your plotting speeds will be bottlenecked by that 1070ti, so just one 2630L v2 will be good enough, and you will not need to worry about NUMA.
If you switch to 3060 ti, the RAM top speed will be limited to 1,600, and the bus bandwidth will drop from ~59 to 51 or so what will limit the raw plotting speed. However, that CPU will start suffering as the final xfr to HDs will also take some toll on it. Basically, you should expect ~20-30 sec extra per plot.
I also thought the 1080ti and the A2000 numbers don’t make sense, but I am running out of ideas, I am getting a 3060ti in the next couple of days and will see if it will work out better.
Dumb question, are you running the latest version?
If a DIMM slot is for 2nd CPU you need that CPU
Pretty sure it’s your -t SSD causing the slowdown, not CPU or RAM.
With my 970 PRO, I see maybe 1 or 2 sec slowdown when copying plots to a final destination.
You need to look into LDDR3 RAM. If your mobo supports those, you can use 32 GB modules. I purchased 256 GB 1,866 in 8x 32 GB sticks for ~$130 on eBay. You really want all that 256 GB sitting on just one CPU. Those 2630L chips only take 1,600 RAM, though.
If you remove one CPU, all the RAM and PCIe slots that sits on that CPU are gone, but the box should be fine running with just one. You also save on power draw, as otherwise that other CPU basically sits idle.
Although, what I wanted to do, but at the end failed is really run 2 CPUs, NUMA bind 256 GB RAM and GPU to CPU 1, and see whether the second CPU could pick up the final xfr. Maybe somewhat difficult with just a plotter, but potentially could work with plotter + plot-sink combo (try to NUMA bind the plot-sink to the second CPU, …). That plot-sink is a hidden gem in those binaries. If one could properly set it up with NUMA binding, potentially the lower power CPU could be used, but I didn’t see those on eBay, or at least delivery time was long (coming from China).
I honestly don’t know how people find those deals on Ebay (or find anything on Ebay for that matter, what a mess that site)
I searched all over and the only reasonable priced 256GB I could find was 16x16 in the US, that’s the whole reason I went this way.
+130 + shipping + import tax, about 180,- euro, still (much) cheaper than buying local though.
Anyway, I only have 250 TB to plot, so not looking to break records for plotting.
The system comes in at just under $500 including the 1070ti and I’m sure I can sell it for that or more a few months down the road
edit: I will report back when the Numa problems start hitting haha